Monte Carlo Methods for Partial Differential Equations Prof. Michael Mascagni

advertisement
MCM for PDEs
Monte Carlo Methods for
Partial Differential Equations
Prof. Michael Mascagni
Department of Computer Science
Department of Mathematics
Department of Scientific Computing
Florida State University, Tallahassee, FL 32306 USA
E-mail: mascagni@fsu.edu or mascagni@math.ethz.ch
URL: http://www.cs.fsu.edu/∼mascagni
In collaboration with Drs. Marcia O. Fenley, and Nikolai Simonov and Messrs. Alexander
Silalahi, and James McClain
Research supported by ARO, DOE/ASCI, NATO, and NSF
MCM for PDEs
Introduction
Early History of MCMs for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches of Reaction-Diffusion Equations
Monte Carlo Methods for PDEs from Fluid Mechanics
Probabilistic Representations for Other PDEs
Monte Carlo Methods and Linear Algebra
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Parallel N-body Potential Evaluation
Bibliography
MCM for PDEs
Introduction
Early History of MCMs for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches of Reaction-Diffusion Equations
Monte Carlo Methods for PDEs from Fluid Mechanics
Probabilistic Representations for Other PDEs
Monte Carlo Methods and Linear Algebra
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Parallel N-body Potential Evaluation
Bibliography
MCM for PDEs
Introduction
Early History of MCMs for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches of Reaction-Diffusion Equations
Monte Carlo Methods for PDEs from Fluid Mechanics
Probabilistic Representations for Other PDEs
Monte Carlo Methods and Linear Algebra
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Parallel N-body Potential Evaluation
Bibliography
MCM for PDEs
Introduction
Early History of MCMs for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches of Reaction-Diffusion Equations
Monte Carlo Methods for PDEs from Fluid Mechanics
Probabilistic Representations for Other PDEs
Monte Carlo Methods and Linear Algebra
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Parallel N-body Potential Evaluation
Bibliography
MCM for PDEs
Introduction
Early History of MCMs for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches of Reaction-Diffusion Equations
Monte Carlo Methods for PDEs from Fluid Mechanics
Probabilistic Representations for Other PDEs
Monte Carlo Methods and Linear Algebra
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Parallel N-body Potential Evaluation
Bibliography
MCM for PDEs
Early History of MCMs for PDEs
Early History of MCMs for PDEs
1. Courant, Friedrichs, and Lewy: Their pivotal 1928 paper has
probabilistic interpretations and MC algorithms for linear elliptic
and parabolic problems
2. Fermi/Ulam/von Neumann: Atomic bomb calculations were done
using Monte Carlo methods for neutron transport, their success
inspired much post-War work especially in nuclear reactor design
3. Kac and Donsker: Used large deviation calculations to estimate
eigenvalues of a linear Schrödinger equation
4. Forsythe and Leibler: Derived a MCM for solving special linear
systems related to discrete elliptic PDE problems
MCM for PDEs
Early History of MCMs for PDEs
Early History of MCMs for PDEs
1. Courant, Friedrichs, and Lewy: Their pivotal 1928 paper has
probabilistic interpretations and MC algorithms for linear elliptic
and parabolic problems
2. Fermi/Ulam/von Neumann: Atomic bomb calculations were done
using Monte Carlo methods for neutron transport, their success
inspired much post-War work especially in nuclear reactor design
3. Kac and Donsker: Used large deviation calculations to estimate
eigenvalues of a linear Schrödinger equation
4. Forsythe and Leibler: Derived a MCM for solving special linear
systems related to discrete elliptic PDE problems
MCM for PDEs
Early History of MCMs for PDEs
Early History of MCMs for PDEs
1. Courant, Friedrichs, and Lewy: Their pivotal 1928 paper has
probabilistic interpretations and MC algorithms for linear elliptic
and parabolic problems
2. Fermi/Ulam/von Neumann: Atomic bomb calculations were done
using Monte Carlo methods for neutron transport, their success
inspired much post-War work especially in nuclear reactor design
3. Kac and Donsker: Used large deviation calculations to estimate
eigenvalues of a linear Schrödinger equation
4. Forsythe and Leibler: Derived a MCM for solving special linear
systems related to discrete elliptic PDE problems
MCM for PDEs
Early History of MCMs for PDEs
Early History of MCMs for PDEs
1. Courant, Friedrichs, and Lewy: Their pivotal 1928 paper has
probabilistic interpretations and MC algorithms for linear elliptic
and parabolic problems
2. Fermi/Ulam/von Neumann: Atomic bomb calculations were done
using Monte Carlo methods for neutron transport, their success
inspired much post-War work especially in nuclear reactor design
3. Kac and Donsker: Used large deviation calculations to estimate
eigenvalues of a linear Schrödinger equation
4. Forsythe and Leibler: Derived a MCM for solving special linear
systems related to discrete elliptic PDE problems
MCM for PDEs
Early History of MCMs for PDEs
Early History of MCMs for PDEs
1. Curtiss: Compared Monte Carlo, direct and iterative solution
methods for Ax = b
I
I
General conclusions of all this work (as other methods were
explored) is that random walk methods do worse than conventional
methods on serial computers except when modest precision and
few solution values are required
Much of this “conventional wisdom” needs revision due to
complexity differences with parallel implementations
MCM for PDEs
Early History of MCMs for PDEs
Early History of MCMs for PDEs
1. Curtiss: Compared Monte Carlo, direct and iterative solution
methods for Ax = b
I
I
General conclusions of all this work (as other methods were
explored) is that random walk methods do worse than conventional
methods on serial computers except when modest precision and
few solution values are required
Much of this “conventional wisdom” needs revision due to
complexity differences with parallel implementations
MCM for PDEs
Early History of MCMs for PDEs
Early History of MCMs for PDEs
1. Curtiss: Compared Monte Carlo, direct and iterative solution
methods for Ax = b
I
I
General conclusions of all this work (as other methods were
explored) is that random walk methods do worse than conventional
methods on serial computers except when modest precision and
few solution values are required
Much of this “conventional wisdom” needs revision due to
complexity differences with parallel implementations
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Elliptic PDEs as Boundary Value Problems
1. Elliptic PDEs describe equilibrium, like the electrostatic field set
up by a charge distribution, or the strain in a beam due to loading
2. No time dependence in elliptic problems so it is natural to have
the interior configuration satisfy a PDE with boundary conditions
to choose a particular global solution
3. Elliptic PDEs are thus part of boundary value problems (BVPs)
such as the famous Dirichlet problem for Laplace’s equation:
1
∆u(x) = 0, x ∈ Ω, u(x) = g(x), x ∈ ∂Ω
2
(1)
4. Here Ω ⊂ Rs is a open set (domain) with a smooth boundary ∂Ω
and g(x) is the given boundary condition
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Elliptic PDEs as Boundary Value Problems
1. Elliptic PDEs describe equilibrium, like the electrostatic field set
up by a charge distribution, or the strain in a beam due to loading
2. No time dependence in elliptic problems so it is natural to have
the interior configuration satisfy a PDE with boundary conditions
to choose a particular global solution
3. Elliptic PDEs are thus part of boundary value problems (BVPs)
such as the famous Dirichlet problem for Laplace’s equation:
1
∆u(x) = 0, x ∈ Ω, u(x) = g(x), x ∈ ∂Ω
2
(1)
4. Here Ω ⊂ Rs is a open set (domain) with a smooth boundary ∂Ω
and g(x) is the given boundary condition
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Elliptic PDEs as Boundary Value Problems
1. Elliptic PDEs describe equilibrium, like the electrostatic field set
up by a charge distribution, or the strain in a beam due to loading
2. No time dependence in elliptic problems so it is natural to have
the interior configuration satisfy a PDE with boundary conditions
to choose a particular global solution
3. Elliptic PDEs are thus part of boundary value problems (BVPs)
such as the famous Dirichlet problem for Laplace’s equation:
1
∆u(x) = 0, x ∈ Ω, u(x) = g(x), x ∈ ∂Ω
2
(1)
4. Here Ω ⊂ Rs is a open set (domain) with a smooth boundary ∂Ω
and g(x) is the given boundary condition
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Elliptic PDEs as Boundary Value Problems
1. Elliptic PDEs describe equilibrium, like the electrostatic field set
up by a charge distribution, or the strain in a beam due to loading
2. No time dependence in elliptic problems so it is natural to have
the interior configuration satisfy a PDE with boundary conditions
to choose a particular global solution
3. Elliptic PDEs are thus part of boundary value problems (BVPs)
such as the famous Dirichlet problem for Laplace’s equation:
1
∆u(x) = 0, x ∈ Ω, u(x) = g(x), x ∈ ∂Ω
2
(1)
4. Here Ω ⊂ Rs is a open set (domain) with a smooth boundary ∂Ω
and g(x) is the given boundary condition
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Elliptic PDEs as Boundary Value Problems
I
An important equivalence for the Laplace equation is the mean
value property (MVP), i.e. if u(x) is a solution to (1) then:
Z
1
u(y ) dy
u(x) =
|∂S n (x, r )| ∂S n (x,r )
∂S n (x, r ) is the surface of an n-dimensional sphere centered at x
with radius r
I
AnotherRway to express u(x) is via the Green’s function:
u(x) = ∂Ω G(x, y )u(y ) dy
I
Showing a function has the MVP and the right boundary values
establishes it as the unique solution to (1)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Elliptic PDEs as Boundary Value Problems
I
An important equivalence for the Laplace equation is the mean
value property (MVP), i.e. if u(x) is a solution to (1) then:
Z
1
u(y ) dy
u(x) =
|∂S n (x, r )| ∂S n (x,r )
∂S n (x, r ) is the surface of an n-dimensional sphere centered at x
with radius r
I
AnotherRway to express u(x) is via the Green’s function:
u(x) = ∂Ω G(x, y )u(y ) dy
I
Showing a function has the MVP and the right boundary values
establishes it as the unique solution to (1)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Elliptic PDEs as Boundary Value Problems
I
An important equivalence for the Laplace equation is the mean
value property (MVP), i.e. if u(x) is a solution to (1) then:
Z
1
u(y ) dy
u(x) =
|∂S n (x, r )| ∂S n (x,r )
∂S n (x, r ) is the surface of an n-dimensional sphere centered at x
with radius r
I
AnotherRway to express u(x) is via the Green’s function:
u(x) = ∂Ω G(x, y )u(y ) dy
I
Showing a function has the MVP and the right boundary values
establishes it as the unique solution to (1)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Early this century probabilists placed measures on different sets
including sets of continuous functions
1.
2.
3.
4.
I
Called Wiener measure
x2
1
Gaussian based: √2πt
e− 2t
Sample paths are Brownian motion
Related to linear PDEs
E.g. u(x) = Ex [g(β(τ∂Ω ))] is the Wiener integral representation of
the solution to (1), to prove it we must check:
1. u(x) = g(x) on ∂Ω
2. u(x) has the MVP
I
Interpretation via Brownian motion and/or a probabilistic Green’s
function
I
Important: τ∂Ω = first passage (hitting) time of the path β(·)
started at x to ∂Ω, statistics based on this random variable are
intimately related to elliptic problems
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Can generalize Wiener integrals to different BVPs via the
relationship between elliptic operators, stochastic differential
equations (SDEs), and the Feynman-Kac formula
I
E.g. consider the general elliptic PDE:
Lu(x) − c(x)u(x) = f (x), x ∈ Ω, c(x) ≥ 0,
u(x) = g(x), x ∈ ∂Ω
(2.1a)
where L is an elliptic partial differential operator of the form:
L=
s
s
X
∂
1X
∂2
+
bi (x)
,
aij (x)
2
∂xi ∂xj
∂xi
i,j=1
i=1
(2.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
Can generalize Wiener integrals to different BVPs via the
relationship between elliptic operators, stochastic differential
equations (SDEs), and the Feynman-Kac formula
I
E.g. consider the general elliptic PDE:
Lu(x) − c(x)u(x) = f (x), x ∈ Ω, c(x) ≥ 0,
u(x) = g(x), x ∈ ∂Ω
(2.1a)
where L is an elliptic partial differential operator of the form:
L=
s
s
X
∂
1X
∂2
+
bi (x)
,
aij (x)
2
∂xi ∂xj
∂xi
i,j=1
i=1
(2.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
The Wiener integral representation is:
Z τ∂Ω R
g(β(τ∂Ω ))
L
− 0t c(β(s)) ds
u(x) = Ex
− f (β(t)) e
dt
τ∂Ω
0
(2.2a)
the expectation is w.r.t. paths which are solutions to the following
(vector) SDE:
dβ(t) = σ(β(t)) dW (t) + b(β(t)) dt, β(0) = x
I
I
(2.2b)
The matrix σ(·) is the Choleski factor (matrix-like square root) of
aij (·) in (2.1b)
To use these ideas to construct MCMs for elliptic BVPs one
must:
1. Simulate sample paths via SDEs (2.2b)
2. Evaluate (2.2a) on the sample paths
3. Sample until variance is acceptable
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
The Wiener integral representation is:
Z τ∂Ω R
g(β(τ∂Ω ))
L
− 0t c(β(s)) ds
u(x) = Ex
− f (β(t)) e
dt
τ∂Ω
0
(2.2a)
the expectation is w.r.t. paths which are solutions to the following
(vector) SDE:
dβ(t) = σ(β(t)) dW (t) + b(β(t)) dt, β(0) = x
I
I
(2.2b)
The matrix σ(·) is the Choleski factor (matrix-like square root) of
aij (·) in (2.1b)
To use these ideas to construct MCMs for elliptic BVPs one
must:
1. Simulate sample paths via SDEs (2.2b)
2. Evaluate (2.2a) on the sample paths
3. Sample until variance is acceptable
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
The Wiener integral representation is:
Z τ∂Ω R
g(β(τ∂Ω ))
L
− 0t c(β(s)) ds
u(x) = Ex
− f (β(t)) e
dt
τ∂Ω
0
(2.2a)
the expectation is w.r.t. paths which are solutions to the following
(vector) SDE:
dβ(t) = σ(β(t)) dW (t) + b(β(t)) dt, β(0) = x
I
I
(2.2b)
The matrix σ(·) is the Choleski factor (matrix-like square root) of
aij (·) in (2.1b)
To use these ideas to construct MCMs for elliptic BVPs one
must:
1. Simulate sample paths via SDEs (2.2b)
2. Evaluate (2.2a) on the sample paths
3. Sample until variance is acceptable
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
The Wiener integral representation is:
Z τ∂Ω R
g(β(τ∂Ω ))
L
− 0t c(β(s)) ds
u(x) = Ex
− f (β(t)) e
dt
τ∂Ω
0
(2.2a)
the expectation is w.r.t. paths which are solutions to the following
(vector) SDE:
dβ(t) = σ(β(t)) dW (t) + b(β(t)) dt, β(0) = x
I
I
(2.2b)
The matrix σ(·) is the Choleski factor (matrix-like square root) of
aij (·) in (2.1b)
To use these ideas to construct MCMs for elliptic BVPs one
must:
1. Simulate sample paths via SDEs (2.2b)
2. Evaluate (2.2a) on the sample paths
3. Sample until variance is acceptable
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
The Wiener integral representation is:
Z τ∂Ω R
g(β(τ∂Ω ))
L
− 0t c(β(s)) ds
u(x) = Ex
− f (β(t)) e
dt
τ∂Ω
0
(2.2a)
the expectation is w.r.t. paths which are solutions to the following
(vector) SDE:
dβ(t) = σ(β(t)) dW (t) + b(β(t)) dt, β(0) = x
I
I
(2.2b)
The matrix σ(·) is the Choleski factor (matrix-like square root) of
aij (·) in (2.1b)
To use these ideas to construct MCMs for elliptic BVPs one
must:
1. Simulate sample paths via SDEs (2.2b)
2. Evaluate (2.2a) on the sample paths
3. Sample until variance is acceptable
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Elliptic PDEs via Feynman-Kac
Probabilistic Approaches to Elliptic PDEs
I
The Wiener integral representation is:
Z τ∂Ω R
g(β(τ∂Ω ))
L
− 0t c(β(s)) ds
u(x) = Ex
− f (β(t)) e
dt
τ∂Ω
0
(2.2a)
the expectation is w.r.t. paths which are solutions to the following
(vector) SDE:
dβ(t) = σ(β(t)) dW (t) + b(β(t)) dt, β(0) = x
I
I
(2.2b)
The matrix σ(·) is the Choleski factor (matrix-like square root) of
aij (·) in (2.1b)
To use these ideas to construct MCMs for elliptic BVPs one
must:
1. Simulate sample paths via SDEs (2.2b)
2. Evaluate (2.2a) on the sample paths
3. Sample until variance is acceptable
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches to Parabolic PDEs via
Feynman-Kac
I
Can generalize Wiener integrals to a wide class of IBVPs via the
relationship between elliptic operators, stochastic differential
equations (SDEs), and the Feynman-Kac formula
I
Recall that t → ∞ parabolic → elliptic
I
E.g. consider the general elliptic PDE:
ut = Lu(x) − c(x)u(x) − f (x), x ∈ Ω, c(x) ≥ 0,
u(x) = g(x), x ∈ ∂Ω
(2.3a)
where L is an elliptic partial differential operator of the form:
L=
s
s
X
1X
∂2
∂
+
bi (x)
,
aij (x)
2
∂xi ∂xj
∂xi
i,j=1
i=1
(2.3b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches to Parabolic PDEs via
Feynman-Kac
I
Can generalize Wiener integrals to a wide class of IBVPs via the
relationship between elliptic operators, stochastic differential
equations (SDEs), and the Feynman-Kac formula
I
Recall that t → ∞ parabolic → elliptic
I
E.g. consider the general elliptic PDE:
ut = Lu(x) − c(x)u(x) − f (x), x ∈ Ω, c(x) ≥ 0,
u(x) = g(x), x ∈ ∂Ω
(2.3a)
where L is an elliptic partial differential operator of the form:
L=
s
s
X
1X
∂2
∂
+
bi (x)
,
aij (x)
2
∂xi ∂xj
∂xi
i,j=1
i=1
(2.3b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches to Parabolic PDEs via
Feynman-Kac
I
Can generalize Wiener integrals to a wide class of IBVPs via the
relationship between elliptic operators, stochastic differential
equations (SDEs), and the Feynman-Kac formula
I
Recall that t → ∞ parabolic → elliptic
I
E.g. consider the general elliptic PDE:
ut = Lu(x) − c(x)u(x) − f (x), x ∈ Ω, c(x) ≥ 0,
u(x) = g(x), x ∈ ∂Ω
(2.3a)
where L is an elliptic partial differential operator of the form:
L=
s
s
X
1X
∂2
∂
+
bi (x)
,
aij (x)
2
∂xi ∂xj
∂xi
i,j=1
i=1
(2.3b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches to Parabolic PDEs via
Feynman-Kac
I
The Wiener integral representation is:
u(x, t) =
ExL
Z
g(β(τ∂Ω ) −
t
f (β(t))e
−
Rt
0
c(β(s)) ds
dt
(2.4a)
0
the expectation is w.r.t. paths which are solutions to the following
(vector) SDE:
dβ(t) = σ(β(t)) dW (t) + b(β(t)) dt, β(0) = x
I
(2.4b)
The matrix σ(·) is the Choleski factor (matrix-like square root) of
aij (·) in (2.3b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Probabilistic Approaches to Parabolic PDEs via
Feynman-Kac
I
The Wiener integral representation is:
u(x, t) =
ExL
Z
g(β(τ∂Ω ) −
t
f (β(t))e
−
Rt
0
c(β(s)) ds
dt
(2.4a)
0
the expectation is w.r.t. paths which are solutions to the following
(vector) SDE:
dβ(t) = σ(β(t)) dW (t) + b(β(t)) dt, β(0) = x
I
(2.4b)
The matrix σ(·) is the Choleski factor (matrix-like square root) of
aij (·) in (2.3b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Different SDEs, Different Processes, Different
Equations
I
The SDE gives us a process, and the process defines L (note: a
complete definition of L includes the boundary conditions)
I
We have solved only the Dirichlet problem, what about other
BCs?
I
Neumann Boundary Conditions:
I
If one uses reflecting Brownian motion, can sample over these
paths
I
Mixed Boundary Conditions: α ∂u
∂n + βu = g(x) on ∂Ω
I
Use reflecting Brownian motion and first passage probabilities,
together
∂u
∂n
= g(x) on ∂Ω
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Different SDEs, Different Processes, Different
Equations
I
The SDE gives us a process, and the process defines L (note: a
complete definition of L includes the boundary conditions)
I
We have solved only the Dirichlet problem, what about other
BCs?
I
Neumann Boundary Conditions:
I
If one uses reflecting Brownian motion, can sample over these
paths
I
Mixed Boundary Conditions: α ∂u
∂n + βu = g(x) on ∂Ω
I
Use reflecting Brownian motion and first passage probabilities,
together
∂u
∂n
= g(x) on ∂Ω
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Different SDEs, Different Processes, Different
Equations
I
The SDE gives us a process, and the process defines L (note: a
complete definition of L includes the boundary conditions)
I
We have solved only the Dirichlet problem, what about other
BCs?
I
Neumann Boundary Conditions:
I
If one uses reflecting Brownian motion, can sample over these
paths
I
Mixed Boundary Conditions: α ∂u
∂n + βu = g(x) on ∂Ω
I
Use reflecting Brownian motion and first passage probabilities,
together
∂u
∂n
= g(x) on ∂Ω
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Different SDEs, Different Processes, Different
Equations
I
The SDE gives us a process, and the process defines L (note: a
complete definition of L includes the boundary conditions)
I
We have solved only the Dirichlet problem, what about other
BCs?
I
Neumann Boundary Conditions:
I
If one uses reflecting Brownian motion, can sample over these
paths
I
Mixed Boundary Conditions: α ∂u
∂n + βu = g(x) on ∂Ω
I
Use reflecting Brownian motion and first passage probabilities,
together
∂u
∂n
= g(x) on ∂Ω
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Different SDEs, Different Processes, Different
Equations
I
The SDE gives us a process, and the process defines L (note: a
complete definition of L includes the boundary conditions)
I
We have solved only the Dirichlet problem, what about other
BCs?
I
Neumann Boundary Conditions:
I
If one uses reflecting Brownian motion, can sample over these
paths
I
Mixed Boundary Conditions: α ∂u
∂n + βu = g(x) on ∂Ω
I
Use reflecting Brownian motion and first passage probabilities,
together
∂u
∂n
= g(x) on ∂Ω
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representation of Parabolic PDEs via Feynman-Kac
Different SDEs, Different Processes, Different
Equations
I
The SDE gives us a process, and the process defines L (note: a
complete definition of L includes the boundary conditions)
I
We have solved only the Dirichlet problem, what about other
BCs?
I
Neumann Boundary Conditions:
I
If one uses reflecting Brownian motion, can sample over these
paths
I
Mixed Boundary Conditions: α ∂u
∂n + βu = g(x) on ∂Ω
I
Use reflecting Brownian motion and first passage probabilities,
together
∂u
∂n
= g(x) on ∂Ω
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Parabolic PDEs and Initial Value Problems
I
Parabolic PDEs are evolution equations: the heat equation
specifies how an initial temperature profile evolves with time, a
pure initial value problem (IVP):
1
∂u
= ∆u, u(x, 0) = u0 (x)
∂t
2
(3.1a)
I
As with elliptic PDEs, there are Feynman-Kac formulas for IVPs
and initial-boundary value problems (IBVPs) for parabolic PDEs
I
Instead of this approach try to use the fundamental solution,
x2
1
which has a real probabilistic flavor, √2πt
e− 2t is the fundamental
solution of (3.1a), in the construction of a MCM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Parabolic PDEs and Initial Value Problems
I
Parabolic PDEs are evolution equations: the heat equation
specifies how an initial temperature profile evolves with time, a
pure initial value problem (IVP):
1
∂u
= ∆u, u(x, 0) = u0 (x)
∂t
2
(3.1a)
I
As with elliptic PDEs, there are Feynman-Kac formulas for IVPs
and initial-boundary value problems (IBVPs) for parabolic PDEs
I
Instead of this approach try to use the fundamental solution,
x2
1
e− 2t is the fundamental
which has a real probabilistic flavor, √2πt
solution of (3.1a), in the construction of a MCM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Parabolic PDEs and Initial Value Problems
I
Parabolic PDEs are evolution equations: the heat equation
specifies how an initial temperature profile evolves with time, a
pure initial value problem (IVP):
1
∂u
= ∆u, u(x, 0) = u0 (x)
∂t
2
(3.1a)
I
As with elliptic PDEs, there are Feynman-Kac formulas for IVPs
and initial-boundary value problems (IBVPs) for parabolic PDEs
I
Instead of this approach try to use the fundamental solution,
x2
1
e− 2t is the fundamental
which has a real probabilistic flavor, √2πt
solution of (3.1a), in the construction of a MCM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the IVP in (3.1a), if u0 (x) = δ(x − x0 ) (spike at x = x0 ),
(x−x0 )2
1
the exact solution is u(x, t) = √2πt
e− 2t , can interpret this as
u(x, t) is N(x0 , t) for MCM sampling of values of u(x, t)
I
To solve (3.1a) with u0 (x) general, must approximate u0 (x) with
spikes and “move” the spikes via their individual normal
distributions
I
The approximation of a smooth u0 (x) by spikes is quite poor, and
so the MCM above gives a solution with large statistical
fluctuations (variance)
I
Instead, can solve for the gradient of u(x, t) and integrate back to
give a better solution, i.e. if we call v (x, t) = ∂u(x,t)
∂x , v (x, t) solves:
∂v
1
∂u0 (x)
= ∆v , v (x, 0) =
∂t
2
∂x
(3.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the IVP in (3.1a), if u0 (x) = δ(x − x0 ) (spike at x = x0 ),
(x−x0 )2
1
the exact solution is u(x, t) = √2πt
e− 2t , can interpret this as
u(x, t) is N(x0 , t) for MCM sampling of values of u(x, t)
I
To solve (3.1a) with u0 (x) general, must approximate u0 (x) with
spikes and “move” the spikes via their individual normal
distributions
I
The approximation of a smooth u0 (x) by spikes is quite poor, and
so the MCM above gives a solution with large statistical
fluctuations (variance)
I
Instead, can solve for the gradient of u(x, t) and integrate back to
give a better solution, i.e. if we call v (x, t) = ∂u(x,t)
∂x , v (x, t) solves:
∂v
1
∂u0 (x)
= ∆v , v (x, 0) =
∂t
2
∂x
(3.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the IVP in (3.1a), if u0 (x) = δ(x − x0 ) (spike at x = x0 ),
(x−x0 )2
1
the exact solution is u(x, t) = √2πt
e− 2t , can interpret this as
u(x, t) is N(x0 , t) for MCM sampling of values of u(x, t)
I
To solve (3.1a) with u0 (x) general, must approximate u0 (x) with
spikes and “move” the spikes via their individual normal
distributions
I
The approximation of a smooth u0 (x) by spikes is quite poor, and
so the MCM above gives a solution with large statistical
fluctuations (variance)
I
Instead, can solve for the gradient of u(x, t) and integrate back to
give a better solution, i.e. if we call v (x, t) = ∂u(x,t)
∂x , v (x, t) solves:
∂v
1
∂u0 (x)
= ∆v , v (x, 0) =
∂t
2
∂x
(3.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the IVP in (3.1a), if u0 (x) = δ(x − x0 ) (spike at x = x0 ),
(x−x0 )2
1
the exact solution is u(x, t) = √2πt
e− 2t , can interpret this as
u(x, t) is N(x0 , t) for MCM sampling of values of u(x, t)
I
To solve (3.1a) with u0 (x) general, must approximate u0 (x) with
spikes and “move” the spikes via their individual normal
distributions
I
The approximation of a smooth u0 (x) by spikes is quite poor, and
so the MCM above gives a solution with large statistical
fluctuations (variance)
I
Instead, can solve for the gradient of u(x, t) and integrate back to
give a better solution, i.e. if we call v (x, t) = ∂u(x,t)
∂x , v (x, t) solves:
∂v
1
∂u0 (x)
= ∆v , v (x, 0) =
∂t
2
∂x
(3.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
This variance reduction idea is the basis of the random gradient
method:
1.
2.
3.
4.
I
I
I
I
Set up the gradient problem
Initial gradient is spiky
Evolve the gradient via MCM
Integrate to recover function
Rx
u(x, t) = −∞ v (y , t) dy
PN
Note that if v (x, t) = N1 i=1 δ(x − xi ) then
P
u(x, t) = N1 {i|xi ≤x} 1, i.e. a step function
PN
More generally
if v (x, t) = i=1 ai δ(x − xi ) then
P
u(x, t) = {i|xi ≤x} ai , i.e. a step function, here we can
approximate more than monotone initial conditions with ai ’s
negative
Since v (x, t) =
∂u(x,t)
∂x ,
The random gradient method is very efficient and allows solving
pure IVPs on infinite domains without difficulty
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related linear IVP (c is constant):
∂u
1
= ∆u + cu, u(x, 0) = u0 (x)
∂t
2
(3.2a)
I
The first term on the r.h.s. is diffusion and its effect may be
sampled via a normally distributed random number
I
The second term on the r.h.s. is an exponential growth/shrinkage
term, can also sample its effect probabilistically:
Think of dispatching random walkers to do the sampling
I
1.
2.
3.
4.
5.
Choose ∆t s.t. ∆t|c| < 1
Move all walkers via N(0, ∆t)
Create/destroy with prob. = ∆t|c|
c > 0: create by doubling
c < 0: destroy by removal
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for Linear Parabolic IVPs
I
Consider the related gradient IVP:
1
∂u0 (x)
∂v
= ∆v + cv , v (x, 0) =
∂t
2
∂x
I
(3.2b)
Let us summarize the algorithm for the MCM to advance the
solution of (3.2a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t|c| < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
If c > 0 create new walkers at those xi where ξi < ∆tc with ξi
U[0, 1]
5. If c < 0 destroy walkers at those xi where ξi P
< −∆tc with ξi U[0, 1]
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the linear PDE version of the random gradient method
(RGM)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Consider the IVP for a nonlinear scalar reaction diffusion
equation:
∂u
1
= ∆u + c(u), u(x, 0) = u0 (x)
∂t
2
I
The associated gradient equation is:
∂v
1
∂u0 (x)
= ∆v + c 0 (u)v , v (x, 0) =
∂t
2
∂x
I
(3.3a)
(3.3b)
The similarity of (3.3b) to (3.2b) make it clear how to extend the
RGM method to these nonlinear scalar reaction diffusion
equations
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Consider the IVP for a nonlinear scalar reaction diffusion
equation:
∂u
1
= ∆u + c(u), u(x, 0) = u0 (x)
∂t
2
I
The associated gradient equation is:
∂v
1
∂u0 (x)
= ∆v + c 0 (u)v , v (x, 0) =
∂t
2
∂x
I
(3.3a)
(3.3b)
The similarity of (3.3b) to (3.2b) make it clear how to extend the
RGM method to these nonlinear scalar reaction diffusion
equations
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Consider the IVP for a nonlinear scalar reaction diffusion
equation:
∂u
1
= ∆u + c(u), u(x, 0) = u0 (x)
∂t
2
I
The associated gradient equation is:
∂v
1
∂u0 (x)
= ∆v + c 0 (u)v , v (x, 0) =
∂t
2
∂x
I
(3.3a)
(3.3b)
The similarity of (3.3b) to (3.2b) make it clear how to extend the
RGM method to these nonlinear scalar reaction diffusion
equations
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Summary of the algorithm for the RGM to advance the solution of
(3.3a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t (supu |c 0 (u)|) < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
At those xi where c 0 (u) > 0 create new walkers if ξi < ∆tc 0 (u) with
ξi U[0, 1]
5. At those xi where c 0 (u) < 0 destroy walkers if ξi < −∆tc 0 (u) with ξi
U[0, 1]
P
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the nonlinear scalar reaction diffusion version of the RGM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Summary of the algorithm for the RGM to advance the solution of
(3.3a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t (supu |c 0 (u)|) < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
At those xi where c 0 (u) > 0 create new walkers if ξi < ∆tc 0 (u) with
ξi U[0, 1]
5. At those xi where c 0 (u) < 0 destroy walkers if ξi < −∆tc 0 (u) with ξi
U[0, 1]
P
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the nonlinear scalar reaction diffusion version of the RGM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Summary of the algorithm for the RGM to advance the solution of
(3.3a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t (supu |c 0 (u)|) < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
At those xi where c 0 (u) > 0 create new walkers if ξi < ∆tc 0 (u) with
ξi U[0, 1]
5. At those xi where c 0 (u) < 0 destroy walkers if ξi < −∆tc 0 (u) with ξi
U[0, 1]
P
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the nonlinear scalar reaction diffusion version of the RGM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Summary of the algorithm for the RGM to advance the solution of
(3.3a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t (supu |c 0 (u)|) < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
At those xi where c 0 (u) > 0 create new walkers if ξi < ∆tc 0 (u) with
ξi U[0, 1]
5. At those xi where c 0 (u) < 0 destroy walkers if ξi < −∆tc 0 (u) with ξi
U[0, 1]
P
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the nonlinear scalar reaction diffusion version of the RGM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Summary of the algorithm for the RGM to advance the solution of
(3.3a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t (supu |c 0 (u)|) < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
At those xi where c 0 (u) > 0 create new walkers if ξi < ∆tc 0 (u) with
ξi U[0, 1]
5. At those xi where c 0 (u) < 0 destroy walkers if ξi < −∆tc 0 (u) with ξi
U[0, 1]
P
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the nonlinear scalar reaction diffusion version of the RGM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Summary of the algorithm for the RGM to advance the solution of
(3.3a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t (supu |c 0 (u)|) < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
At those xi where c 0 (u) > 0 create new walkers if ξi < ∆tc 0 (u) with
ξi U[0, 1]
5. At those xi where c 0 (u) < 0 destroy walkers if ξi < −∆tc 0 (u) with ξi
U[0, 1]
P
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the nonlinear scalar reaction diffusion version of the RGM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Summary of the algorithm for the RGM to advance the solution of
(3.3a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t (supu |c 0 (u)|) < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
At those xi where c 0 (u) > 0 create new walkers if ξi < ∆tc 0 (u) with
ξi U[0, 1]
5. At those xi where c 0 (u) < 0 destroy walkers if ξi < −∆tc 0 (u) with ξi
U[0, 1]
P
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the nonlinear scalar reaction diffusion version of the RGM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Summary of the algorithm for the RGM to advance the solution of
(3.3a) one time step using ∆t:
P
Represent v (x, t) = N1 Ni=1 δ(x − xi )
Choose ∆t s.t. ∆t (supu |c 0 (u)|) < 1
Move xi to xi + ηi where ηi is N(0, ∆t)
At those xi where c 0 (u) > 0 create new walkers if ξi < ∆tc 0 (u) with
ξi U[0, 1]
5. At those xi where c 0 (u) < 0 destroy walkers if ξi < −∆tc 0 (u) with ξi
U[0, 1]
P
6. Over the remaining points, u(x, t + ∆t) = N1 {i|xi ≤x} 1
1.
2.
3.
4.
I
This is the nonlinear scalar reaction diffusion version of the RGM
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Differences/advantages of the RGM and conventional
(finite-difference/finite-element) methods:
1. It is computationally easy to sample from N(0, ∆t)
2. Only costly operation each time step is to sort the remaining cohort
of walkers by their position
3. Can use ensemble averaging to reduce the variance of the solution
4. Can choose to use either gradient “particles” of equal mass or
allow the mass of each particle to change
5. The RGM is adaptive, computational elements (pieces of the
gradient) are created where c 0 (u) is greatest, this is where sharp
fronts appear and so fewer total computational elements are
needed to spatially resolve jump-like solutions
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Differences/advantages of the RGM and conventional
(finite-difference/finite-element) methods:
1. It is computationally easy to sample from N(0, ∆t)
2. Only costly operation each time step is to sort the remaining cohort
of walkers by their position
3. Can use ensemble averaging to reduce the variance of the solution
4. Can choose to use either gradient “particles” of equal mass or
allow the mass of each particle to change
5. The RGM is adaptive, computational elements (pieces of the
gradient) are created where c 0 (u) is greatest, this is where sharp
fronts appear and so fewer total computational elements are
needed to spatially resolve jump-like solutions
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Differences/advantages of the RGM and conventional
(finite-difference/finite-element) methods:
1. It is computationally easy to sample from N(0, ∆t)
2. Only costly operation each time step is to sort the remaining cohort
of walkers by their position
3. Can use ensemble averaging to reduce the variance of the solution
4. Can choose to use either gradient “particles” of equal mass or
allow the mass of each particle to change
5. The RGM is adaptive, computational elements (pieces of the
gradient) are created where c 0 (u) is greatest, this is where sharp
fronts appear and so fewer total computational elements are
needed to spatially resolve jump-like solutions
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Differences/advantages of the RGM and conventional
(finite-difference/finite-element) methods:
1. It is computationally easy to sample from N(0, ∆t)
2. Only costly operation each time step is to sort the remaining cohort
of walkers by their position
3. Can use ensemble averaging to reduce the variance of the solution
4. Can choose to use either gradient “particles” of equal mass or
allow the mass of each particle to change
5. The RGM is adaptive, computational elements (pieces of the
gradient) are created where c 0 (u) is greatest, this is where sharp
fronts appear and so fewer total computational elements are
needed to spatially resolve jump-like solutions
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Differences/advantages of the RGM and conventional
(finite-difference/finite-element) methods:
1. It is computationally easy to sample from N(0, ∆t)
2. Only costly operation each time step is to sort the remaining cohort
of walkers by their position
3. Can use ensemble averaging to reduce the variance of the solution
4. Can choose to use either gradient “particles” of equal mass or
allow the mass of each particle to change
5. The RGM is adaptive, computational elements (pieces of the
gradient) are created where c 0 (u) is greatest, this is where sharp
fronts appear and so fewer total computational elements are
needed to spatially resolve jump-like solutions
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM for Nonlinear Parabolic IVPs
I
Differences/advantages of the RGM and conventional
(finite-difference/finite-element) methods:
1. It is computationally easy to sample from N(0, ∆t)
2. Only costly operation each time step is to sort the remaining cohort
of walkers by their position
3. Can use ensemble averaging to reduce the variance of the solution
4. Can choose to use either gradient “particles” of equal mass or
allow the mass of each particle to change
5. The RGM is adaptive, computational elements (pieces of the
gradient) are created where c 0 (u) is greatest, this is where sharp
fronts appear and so fewer total computational elements are
needed to spatially resolve jump-like solutions
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM in 2-Dimensions
I
The RGM in 2D is the same as in 1D except for recovery from
the gradient
I
Write u(x) = G(x, y) ∗ ∆u(y) = ∇n G(x, y) ∗ ∇n u(y) (integration
by parts)
I
If ∇n u(y) = δ(y − y0 ) then u(x) = ∇n G(x, y0 ) =
I
Thus gradient recovery can be done via 2 n-body evaluations
with charges n1 and n2 or with 1 Hilbert matrix application to the
complex vector with n
I
Can (and do) use the Rokhlin-Greengard fast multipole algorithm
for gradient recovery
I
Initial gradient distribution comes from a detailed contour plot
−1 (x−y0 )·n
2π kx−y0 k2
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM in 2-Dimensions
I
The RGM in 2D is the same as in 1D except for recovery from
the gradient
I
Write u(x) = G(x, y) ∗ ∆u(y) = ∇n G(x, y) ∗ ∇n u(y) (integration
by parts)
I
If ∇n u(y) = δ(y − y0 ) then u(x) = ∇n G(x, y0 ) =
I
Thus gradient recovery can be done via 2 n-body evaluations
with charges n1 and n2 or with 1 Hilbert matrix application to the
complex vector with n
I
Can (and do) use the Rokhlin-Greengard fast multipole algorithm
for gradient recovery
I
Initial gradient distribution comes from a detailed contour plot
−1 (x−y0 )·n
2π kx−y0 k2
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM in 2-Dimensions
I
The RGM in 2D is the same as in 1D except for recovery from
the gradient
I
Write u(x) = G(x, y) ∗ ∆u(y) = ∇n G(x, y) ∗ ∇n u(y) (integration
by parts)
I
If ∇n u(y) = δ(y − y0 ) then u(x) = ∇n G(x, y0 ) =
I
Thus gradient recovery can be done via 2 n-body evaluations
with charges n1 and n2 or with 1 Hilbert matrix application to the
complex vector with n
I
Can (and do) use the Rokhlin-Greengard fast multipole algorithm
for gradient recovery
I
Initial gradient distribution comes from a detailed contour plot
−1 (x−y0 )·n
2π kx−y0 k2
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM in 2-Dimensions
I
The RGM in 2D is the same as in 1D except for recovery from
the gradient
I
Write u(x) = G(x, y) ∗ ∆u(y) = ∇n G(x, y) ∗ ∇n u(y) (integration
by parts)
I
If ∇n u(y) = δ(y − y0 ) then u(x) = ∇n G(x, y0 ) =
I
Thus gradient recovery can be done via 2 n-body evaluations
with charges n1 and n2 or with 1 Hilbert matrix application to the
complex vector with n
I
Can (and do) use the Rokhlin-Greengard fast multipole algorithm
for gradient recovery
I
Initial gradient distribution comes from a detailed contour plot
−1 (x−y0 )·n
2π kx−y0 k2
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM in 2-Dimensions
I
The RGM in 2D is the same as in 1D except for recovery from
the gradient
I
Write u(x) = G(x, y) ∗ ∆u(y) = ∇n G(x, y) ∗ ∇n u(y) (integration
by parts)
I
If ∇n u(y) = δ(y − y0 ) then u(x) = ∇n G(x, y0 ) =
I
Thus gradient recovery can be done via 2 n-body evaluations
with charges n1 and n2 or with 1 Hilbert matrix application to the
complex vector with n
I
Can (and do) use the Rokhlin-Greengard fast multipole algorithm
for gradient recovery
I
Initial gradient distribution comes from a detailed contour plot
−1 (x−y0 )·n
2π kx−y0 k2
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
The RGM in 2-Dimensions
I
The RGM in 2D is the same as in 1D except for recovery from
the gradient
I
Write u(x) = G(x, y) ∗ ∆u(y) = ∇n G(x, y) ∗ ∇n u(y) (integration
by parts)
I
If ∇n u(y) = δ(y − y0 ) then u(x) = ∇n G(x, y0 ) =
I
Thus gradient recovery can be done via 2 n-body evaluations
with charges n1 and n2 or with 1 Hilbert matrix application to the
complex vector with n
I
Can (and do) use the Rokhlin-Greengard fast multipole algorithm
for gradient recovery
I
Initial gradient distribution comes from a detailed contour plot
−1 (x−y0 )·n
2π kx−y0 k2
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Another MCM for a Nonlinear Parabolic PDE from
Fluid Dynamics
I
A model equation for fluid dynamics is Berger’s equation in
one-dimension, as an IVP:
∂u
∂u
∂2u
+u
=
,
∂t
∂x
2 ∂x 2
I
1
φ)
The substitution φ = e− u dx ⇐⇒ u = − ∂(ln
= − φ1 ∂φ
∂x
∂x
converts Berger’s equation to the heat equation (Hopf, 1950):
∂φ
∂2φ
=
,
∂t
2 ∂x 2
I
u(x, 0) = u0 (x)
R
1
φ(x, 0) = e− Rx
0
u0 (ξ) dξ
Using the Feynman-Kac formula√for the IVP for the heat equation
R β(t)
1
u0 (ξ) dξ
one gets that φ(x, t) = Ex [e− 0
], which determines
u(x, t) via the above inversion formula
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Another MCM for a Nonlinear Parabolic PDE from
Fluid Dynamics
I
A model equation for fluid dynamics is Berger’s equation in
one-dimension, as an IVP:
∂u
∂u
∂2u
+u
=
,
∂t
∂x
2 ∂x 2
I
1
φ)
The substitution φ = e− u dx ⇐⇒ u = − ∂(ln
= − φ1 ∂φ
∂x
∂x
converts Berger’s equation to the heat equation (Hopf, 1950):
∂φ
∂2φ
=
,
∂t
2 ∂x 2
I
u(x, 0) = u0 (x)
R
1
φ(x, 0) = e− Rx
0
u0 (ξ) dξ
Using the Feynman-Kac formula√for the IVP for the heat equation
R β(t)
1
u0 (ξ) dξ
one gets that φ(x, t) = Ex [e− 0
], which determines
u(x, t) via the above inversion formula
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Another MCM for a Nonlinear Parabolic PDE from
Fluid Dynamics
I
A model equation for fluid dynamics is Berger’s equation in
one-dimension, as an IVP:
∂u
∂u
∂2u
+u
=
,
∂t
∂x
2 ∂x 2
I
1
φ)
The substitution φ = e− u dx ⇐⇒ u = − ∂(ln
= − φ1 ∂φ
∂x
∂x
converts Berger’s equation to the heat equation (Hopf, 1950):
∂φ
∂2φ
=
,
∂t
2 ∂x 2
I
u(x, 0) = u0 (x)
R
1
φ(x, 0) = e− Rx
0
u0 (ξ) dξ
Using the Feynman-Kac formula√for the IVP for the heat equation
R β(t)
1
u0 (ξ) dξ
one gets that φ(x, t) = Ex [e− 0
], which determines
u(x, t) via the above inversion formula
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for the Schrödinger Equation (Brief)
I
The Schrödinger equation is given by:
−i
~
uτ = ∆u(x) − V (x)u(x)
2π
I
~
Can replace −i 2π
τ = t (imaginary time), to give a linear
parabolic PDE
I
Usually x ∈ R3n where there are n quantum particles, thus we
are in a very high-dimensional case
I
As in the above, use walks, killing, and importance sampling
Interesting variants:
I
1. Diffusion Monte Carlo
2. Greens Function Monte Carlo
3. Path Integral Monte Carlo
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for the Schrödinger Equation (Brief)
I
The Schrödinger equation is given by:
−i
~
uτ = ∆u(x) − V (x)u(x)
2π
I
~
Can replace −i 2π
τ = t (imaginary time), to give a linear
parabolic PDE
I
Usually x ∈ R3n where there are n quantum particles, thus we
are in a very high-dimensional case
I
As in the above, use walks, killing, and importance sampling
Interesting variants:
I
1. Diffusion Monte Carlo
2. Greens Function Monte Carlo
3. Path Integral Monte Carlo
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for the Schrödinger Equation (Brief)
I
The Schrödinger equation is given by:
−i
~
uτ = ∆u(x) − V (x)u(x)
2π
I
~
Can replace −i 2π
τ = t (imaginary time), to give a linear
parabolic PDE
I
Usually x ∈ R3n where there are n quantum particles, thus we
are in a very high-dimensional case
I
As in the above, use walks, killing, and importance sampling
Interesting variants:
I
1. Diffusion Monte Carlo
2. Greens Function Monte Carlo
3. Path Integral Monte Carlo
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for the Schrödinger Equation (Brief)
I
The Schrödinger equation is given by:
−i
~
uτ = ∆u(x) − V (x)u(x)
2π
I
~
Can replace −i 2π
τ = t (imaginary time), to give a linear
parabolic PDE
I
Usually x ∈ R3n where there are n quantum particles, thus we
are in a very high-dimensional case
I
As in the above, use walks, killing, and importance sampling
Interesting variants:
I
1. Diffusion Monte Carlo
2. Greens Function Monte Carlo
3. Path Integral Monte Carlo
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for the Schrödinger Equation (Brief)
I
The Schrödinger equation is given by:
−i
~
uτ = ∆u(x) − V (x)u(x)
2π
I
~
Can replace −i 2π
τ = t (imaginary time), to give a linear
parabolic PDE
I
Usually x ∈ R3n where there are n quantum particles, thus we
are in a very high-dimensional case
I
As in the above, use walks, killing, and importance sampling
Interesting variants:
I
1. Diffusion Monte Carlo
2. Greens Function Monte Carlo
3. Path Integral Monte Carlo
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for the Schrödinger Equation (Brief)
I
The Schrödinger equation is given by:
−i
~
uτ = ∆u(x) − V (x)u(x)
2π
I
~
Can replace −i 2π
τ = t (imaginary time), to give a linear
parabolic PDE
I
Usually x ∈ R3n where there are n quantum particles, thus we
are in a very high-dimensional case
I
As in the above, use walks, killing, and importance sampling
Interesting variants:
I
1. Diffusion Monte Carlo
2. Greens Function Monte Carlo
3. Path Integral Monte Carlo
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for the Schrödinger Equation (Brief)
I
The Schrödinger equation is given by:
−i
~
uτ = ∆u(x) − V (x)u(x)
2π
I
~
Can replace −i 2π
τ = t (imaginary time), to give a linear
parabolic PDE
I
Usually x ∈ R3n where there are n quantum particles, thus we
are in a very high-dimensional case
I
As in the above, use walks, killing, and importance sampling
Interesting variants:
I
1. Diffusion Monte Carlo
2. Greens Function Monte Carlo
3. Path Integral Monte Carlo
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
MCMs for the Schrödinger Equation (Brief)
I
The Schrödinger equation is given by:
−i
~
uτ = ∆u(x) − V (x)u(x)
2π
I
~
Can replace −i 2π
τ = t (imaginary time), to give a linear
parabolic PDE
I
Usually x ∈ R3n where there are n quantum particles, thus we
are in a very high-dimensional case
I
As in the above, use walks, killing, and importance sampling
Interesting variants:
I
1. Diffusion Monte Carlo
2. Greens Function Monte Carlo
3. Path Integral Monte Carlo
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Another MCM for a Nonlinear Parabolic PDE from
Fluid Dynamics
I
I
√
Note when L = 2 ∆ the scaling x → x converts L into the pure
Laplacian
√
Thus can sample from L with β(·) as scaled sample paths
instead of ordinary Brownian motion, this is Brownian scaling
I
Unlike the reaction-diffusion problems solved by the RGM this
equation is an hyperbolic conservation law: ut + (u 2 /2)x = 2 uxx ,
these equations often have jumps that sharpen into
discontinuous shocks as → 0
I
The MCM for Berger’s equation derived above was only possible
because the Hopf-Cole transformation could be used to convert
this problem to the heat equation
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Another MCM for a Nonlinear Parabolic PDE from
Fluid Dynamics
I
I
√
Note when L = 2 ∆ the scaling x → x converts L into the pure
Laplacian
√
Thus can sample from L with β(·) as scaled sample paths
instead of ordinary Brownian motion, this is Brownian scaling
I
Unlike the reaction-diffusion problems solved by the RGM this
equation is an hyperbolic conservation law: ut + (u 2 /2)x = 2 uxx ,
these equations often have jumps that sharpen into
discontinuous shocks as → 0
I
The MCM for Berger’s equation derived above was only possible
because the Hopf-Cole transformation could be used to convert
this problem to the heat equation
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Another MCM for a Nonlinear Parabolic PDE from
Fluid Dynamics
I
I
√
Note when L = 2 ∆ the scaling x → x converts L into the pure
Laplacian
√
Thus can sample from L with β(·) as scaled sample paths
instead of ordinary Brownian motion, this is Brownian scaling
I
Unlike the reaction-diffusion problems solved by the RGM this
equation is an hyperbolic conservation law: ut + (u 2 /2)x = 2 uxx ,
these equations often have jumps that sharpen into
discontinuous shocks as → 0
I
The MCM for Berger’s equation derived above was only possible
because the Hopf-Cole transformation could be used to convert
this problem to the heat equation
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Approaches of Reaction-Diffusion Equations
Another MCM for a Nonlinear Parabolic PDE from
Fluid Dynamics
I
I
√
Note when L = 2 ∆ the scaling x → x converts L into the pure
Laplacian
√
Thus can sample from L with β(·) as scaled sample paths
instead of ordinary Brownian motion, this is Brownian scaling
I
Unlike the reaction-diffusion problems solved by the RGM this
equation is an hyperbolic conservation law: ut + (u 2 /2)x = 2 uxx ,
these equations often have jumps that sharpen into
discontinuous shocks as → 0
I
The MCM for Berger’s equation derived above was only possible
because the Hopf-Cole transformation could be used to convert
this problem to the heat equation
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Two-Dimensional Incompressible Fluid Dynamics
I
The equations of two dimensional incompressible fluid dynamics
are:
∂u
1
+ (u · ∇)u = − ∇p + γ∆u
∂t
ρ
∇ · u = 0,
u = (u, v )T
(4.1a)
1. Inviscid fluid: γ = 0 =⇒ Euler equations
2. Viscous fluid: γ 6= 0 =⇒ Navier-Stokes equations
I
∂ψ T
Since ∇ · u = 0 there is a stream function ψ s.t. u = (− ∂ψ
∂y , ∂x ) ,
with the vorticity ξ = ∇ × u we can rewrite as:
∂ξ
+ (u · ∇)ξ = γ∆ξ,
∂t
∆ψ = −ξ
(4.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Two-Dimensional Incompressible Fluid Dynamics
I
The equations of two dimensional incompressible fluid dynamics
are:
∂u
1
+ (u · ∇)u = − ∇p + γ∆u
∂t
ρ
∇ · u = 0,
u = (u, v )T
(4.1a)
1. Inviscid fluid: γ = 0 =⇒ Euler equations
2. Viscous fluid: γ 6= 0 =⇒ Navier-Stokes equations
I
∂ψ T
Since ∇ · u = 0 there is a stream function ψ s.t. u = (− ∂ψ
∂y , ∂x ) ,
with the vorticity ξ = ∇ × u we can rewrite as:
∂ξ
+ (u · ∇)ξ = γ∆ξ,
∂t
∆ψ = −ξ
(4.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Two-Dimensional Incompressible Fluid Dynamics
I
The equations of two dimensional incompressible fluid dynamics
are:
∂u
1
+ (u · ∇)u = − ∇p + γ∆u
∂t
ρ
∇ · u = 0,
u = (u, v )T
(4.1a)
1. Inviscid fluid: γ = 0 =⇒ Euler equations
2. Viscous fluid: γ 6= 0 =⇒ Navier-Stokes equations
I
∂ψ T
Since ∇ · u = 0 there is a stream function ψ s.t. u = (− ∂ψ
∂y , ∂x ) ,
with the vorticity ξ = ∇ × u we can rewrite as:
∂ξ
+ (u · ∇)ξ = γ∆ξ,
∂t
∆ψ = −ξ
(4.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Two-Dimensional Incompressible Fluid Dynamics
I
The equations of two dimensional incompressible fluid dynamics
are:
∂u
1
+ (u · ∇)u = − ∇p + γ∆u
∂t
ρ
∇ · u = 0,
u = (u, v )T
(4.1a)
1. Inviscid fluid: γ = 0 =⇒ Euler equations
2. Viscous fluid: γ 6= 0 =⇒ Navier-Stokes equations
I
∂ψ T
Since ∇ · u = 0 there is a stream function ψ s.t. u = (− ∂ψ
∂y , ∂x ) ,
with the vorticity ξ = ∇ × u we can rewrite as:
∂ξ
+ (u · ∇)ξ = γ∆ξ,
∂t
∆ψ = −ξ
(4.1b)
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Two-Dimensional Incompressible Fluid Dynamics
I
∂z
Recall the material (or total) derivative of z: Dz
Dt = ∂t + (u · ∇)z,
this is the time rate of change in a quantity that is being advected
in a fluid with velocity u
I
We can rewrite the vorticity above formulation as:
Dξ
= γ∆ξ,
Dt
∆ψ = −ξ
(4.1c)
I
If we neglect boundary conditions, can represent
PN
ξ = n=1 ξi δ(x − xi ) (spike discretization of a gradient), since in
2D the fundamental solution of the Poisson equation is
1
∆−1 δ(x − xi ) = − 2π
log |x − xi | we have
P
N
1
ψ = 2π n=1 ξi log |x − xi |
I
Can prove that if ξ is a sum of delta functions at some time t0 , it
remains so for all time t ≥ t0
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Two-Dimensional Incompressible Fluid Dynamics
I
∂z
Recall the material (or total) derivative of z: Dz
Dt = ∂t + (u · ∇)z,
this is the time rate of change in a quantity that is being advected
in a fluid with velocity u
I
We can rewrite the vorticity above formulation as:
Dξ
= γ∆ξ,
Dt
∆ψ = −ξ
(4.1c)
I
If we neglect boundary conditions, can represent
PN
ξ = n=1 ξi δ(x − xi ) (spike discretization of a gradient), since in
2D the fundamental solution of the Poisson equation is
1
log |x − xi | we have
∆−1 δ(x − xi ) = − 2π
P
N
1
ψ = 2π n=1 ξi log |x − xi |
I
Can prove that if ξ is a sum of delta functions at some time t0 , it
remains so for all time t ≥ t0
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Two-Dimensional Incompressible Fluid Dynamics
I
∂z
Recall the material (or total) derivative of z: Dz
Dt = ∂t + (u · ∇)z,
this is the time rate of change in a quantity that is being advected
in a fluid with velocity u
I
We can rewrite the vorticity above formulation as:
Dξ
= γ∆ξ,
Dt
∆ψ = −ξ
(4.1c)
I
If we neglect boundary conditions, can represent
PN
ξ = n=1 ξi δ(x − xi ) (spike discretization of a gradient), since in
2D the fundamental solution of the Poisson equation is
1
log |x − xi | we have
∆−1 δ(x − xi ) = − 2π
P
N
1
ψ = 2π n=1 ξi log |x − xi |
I
Can prove that if ξ is a sum of delta functions at some time t0 , it
remains so for all time t ≥ t0
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Two-Dimensional Incompressible Fluid Dynamics
I
∂z
Recall the material (or total) derivative of z: Dz
Dt = ∂t + (u · ∇)z,
this is the time rate of change in a quantity that is being advected
in a fluid with velocity u
I
We can rewrite the vorticity above formulation as:
Dξ
= γ∆ξ,
Dt
∆ψ = −ξ
(4.1c)
I
If we neglect boundary conditions, can represent
PN
ξ = n=1 ξi δ(x − xi ) (spike discretization of a gradient), since in
2D the fundamental solution of the Poisson equation is
1
log |x − xi | we have
∆−1 δ(x − xi ) = − 2π
P
N
1
ψ = 2π n=1 ξi log |x − xi |
I
Can prove that if ξ is a sum of delta functions at some time t0 , it
remains so for all time t ≥ t0
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
The Vortex Method for the Incompressible Euler’s
Equation
I
These observations on the vortex form of the Euler equations
help extend ideas of a method first proposed by Chorin (Chorin,
1971):
P
1. Represent ξ(x, t) = Ni=1 ξi δ(x − xi ), and so
P
N
1
ψ = 2π
n=1 ξi log |x − xi |
2. Move each of the vortex “blobs” via the ODEs for the x and y
components of ξi :
dxi
dt
dyi
dt
I
= −
1 X ∂ log |xi − xj |
ξj
, i = 1, . . . , N
2π
∂y
j6=i
=
1 X ∂ log |xi − xj |
ξj
, i = 1, . . . , N
2π
∂x
j6=i
Note, no time step size constraint is a priori imposed but
depends on the choice of numerical ODE scheme
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
The Vortex Method for the Incompressible Euler’s
Equation
I
These observations on the vortex form of the Euler equations
help extend ideas of a method first proposed by Chorin (Chorin,
1971):
P
1. Represent ξ(x, t) = Ni=1 ξi δ(x − xi ), and so
P
N
1
ψ = 2π
n=1 ξi log |x − xi |
2. Move each of the vortex “blobs” via the ODEs for the x and y
components of ξi :
dxi
dt
dyi
dt
I
= −
1 X ∂ log |xi − xj |
ξj
, i = 1, . . . , N
2π
∂y
j6=i
=
1 X ∂ log |xi − xj |
ξj
, i = 1, . . . , N
2π
∂x
j6=i
Note, no time step size constraint is a priori imposed but
depends on the choice of numerical ODE scheme
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
The Vortex Method for the Incompressible Euler’s
Equation
I
These observations on the vortex form of the Euler equations
help extend ideas of a method first proposed by Chorin (Chorin,
1971):
P
1. Represent ξ(x, t) = Ni=1 ξi δ(x − xi ), and so
P
N
1
ψ = 2π
n=1 ξi log |x − xi |
2. Move each of the vortex “blobs” via the ODEs for the x and y
components of ξi :
dxi
dt
dyi
dt
I
= −
1 X ∂ log |xi − xj |
ξj
, i = 1, . . . , N
2π
∂y
j6=i
=
1 X ∂ log |xi − xj |
ξj
, i = 1, . . . , N
2π
∂x
j6=i
Note, no time step size constraint is a priori imposed but
depends on the choice of numerical ODE scheme
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
The Vortex Method for the Incompressible Euler’s
Equation
I
These observations on the vortex form of the Euler equations
help extend ideas of a method first proposed by Chorin (Chorin,
1971):
P
1. Represent ξ(x, t) = Ni=1 ξi δ(x − xi ), and so
P
N
1
ψ = 2π
n=1 ξi log |x − xi |
2. Move each of the vortex “blobs” via the ODEs for the x and y
components of ξi :
dxi
dt
dyi
dt
I
= −
1 X ∂ log |xi − xj |
ξj
, i = 1, . . . , N
2π
∂y
j6=i
=
1 X ∂ log |xi − xj |
ξj
, i = 1, . . . , N
2π
∂x
j6=i
Note, no time step size constraint is a priori imposed but
depends on the choice of numerical ODE scheme
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
The Vortex Method for the Incompressible Euler’s
Equation
I
This is not a MCM but only a method for converting the 2D Euler
equations (PDEs) to a system of ODEs which are mathematically
equivalent to the N-body problem (gravitation, particle dynamics)
I
It is common to use other functional forms for the vortex “blobs,"
and hence the stream function is changed, in our case the
PN
stream “blob” function is ψ(x) = n=1 ξi ψ 0 (x − xi ) where
1
ψ 0 (x − xi ) = 2π
log |x − xi |
I
Other choices of stream “blob” functions are ψ 0 (x)
1
s.t. ψ 0 (x) ∼ 2π
log(x) for |x| 0, and ψ 0 (x) → 0 as |x| → 0
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
The Vortex Method for the Incompressible Euler’s
Equation
I
This is not a MCM but only a method for converting the 2D Euler
equations (PDEs) to a system of ODEs which are mathematically
equivalent to the N-body problem (gravitation, particle dynamics)
I
It is common to use other functional forms for the vortex “blobs,"
and hence the stream function is changed, in our case the
PN
stream “blob” function is ψ(x) = n=1 ξi ψ 0 (x − xi ) where
1
ψ 0 (x − xi ) = 2π
log |x − xi |
I
Other choices of stream “blob” functions are ψ 0 (x)
1
s.t. ψ 0 (x) ∼ 2π
log(x) for |x| 0, and ψ 0 (x) → 0 as |x| → 0
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
The Vortex Method for the Incompressible Euler’s
Equation
I
This is not a MCM but only a method for converting the 2D Euler
equations (PDEs) to a system of ODEs which are mathematically
equivalent to the N-body problem (gravitation, particle dynamics)
I
It is common to use other functional forms for the vortex “blobs,"
and hence the stream function is changed, in our case the
PN
stream “blob” function is ψ(x) = n=1 ξi ψ 0 (x − xi ) where
1
ψ 0 (x − xi ) = 2π
log |x − xi |
I
Other choices of stream “blob” functions are ψ 0 (x)
1
s.t. ψ 0 (x) ∼ 2π
log(x) for |x| 0, and ψ 0 (x) → 0 as |x| → 0
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s Random Vortex Method for the
Incompressible Navier-Stokes Equations
I
When γ 6= 0, Euler =⇒ Navier-Stokes and (4.1b) can be
thought of as an equation where vorticity is advected via fluid
(the l.h.s. material derivative), and diffuses due to viscosity (the
r.h.s term)
I
As above, diffusion can be sampled by moving diffusing particles
via the Gaussian (fundamental soln. of the diffusion eqn.)
I
Thus we can extend the inviscid random vortex method (RVM) to
Navier-Stokes through a fractional step approach, do the vorticity
advection via the inviscid RVM and then treat the diffusion of
vorticity equation by moving the vortex randomly
I
This addition for diffusion of vorticity make Chorin’s RGM into a
MCM
I
Note that the RGM is also fractional step, but both steps are
MCMs
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s Random Vortex Method for the
Incompressible Navier-Stokes Equations
I
When γ 6= 0, Euler =⇒ Navier-Stokes and (4.1b) can be
thought of as an equation where vorticity is advected via fluid
(the l.h.s. material derivative), and diffuses due to viscosity (the
r.h.s term)
I
As above, diffusion can be sampled by moving diffusing particles
via the Gaussian (fundamental soln. of the diffusion eqn.)
I
Thus we can extend the inviscid random vortex method (RVM) to
Navier-Stokes through a fractional step approach, do the vorticity
advection via the inviscid RVM and then treat the diffusion of
vorticity equation by moving the vortex randomly
I
This addition for diffusion of vorticity make Chorin’s RGM into a
MCM
I
Note that the RGM is also fractional step, but both steps are
MCMs
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s Random Vortex Method for the
Incompressible Navier-Stokes Equations
I
When γ 6= 0, Euler =⇒ Navier-Stokes and (4.1b) can be
thought of as an equation where vorticity is advected via fluid
(the l.h.s. material derivative), and diffuses due to viscosity (the
r.h.s term)
I
As above, diffusion can be sampled by moving diffusing particles
via the Gaussian (fundamental soln. of the diffusion eqn.)
I
Thus we can extend the inviscid random vortex method (RVM) to
Navier-Stokes through a fractional step approach, do the vorticity
advection via the inviscid RVM and then treat the diffusion of
vorticity equation by moving the vortex randomly
I
This addition for diffusion of vorticity make Chorin’s RGM into a
MCM
I
Note that the RGM is also fractional step, but both steps are
MCMs
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s Random Vortex Method for the
Incompressible Navier-Stokes Equations
I
When γ 6= 0, Euler =⇒ Navier-Stokes and (4.1b) can be
thought of as an equation where vorticity is advected via fluid
(the l.h.s. material derivative), and diffuses due to viscosity (the
r.h.s term)
I
As above, diffusion can be sampled by moving diffusing particles
via the Gaussian (fundamental soln. of the diffusion eqn.)
I
Thus we can extend the inviscid random vortex method (RVM) to
Navier-Stokes through a fractional step approach, do the vorticity
advection via the inviscid RVM and then treat the diffusion of
vorticity equation by moving the vortex randomly
I
This addition for diffusion of vorticity make Chorin’s RGM into a
MCM
I
Note that the RGM is also fractional step, but both steps are
MCMs
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s Random Vortex Method for the
Incompressible Navier-Stokes Equations
I
When γ 6= 0, Euler =⇒ Navier-Stokes and (4.1b) can be
thought of as an equation where vorticity is advected via fluid
(the l.h.s. material derivative), and diffuses due to viscosity (the
r.h.s term)
I
As above, diffusion can be sampled by moving diffusing particles
via the Gaussian (fundamental soln. of the diffusion eqn.)
I
Thus we can extend the inviscid random vortex method (RVM) to
Navier-Stokes through a fractional step approach, do the vorticity
advection via the inviscid RVM and then treat the diffusion of
vorticity equation by moving the vortex randomly
I
This addition for diffusion of vorticity make Chorin’s RGM into a
MCM
I
Note that the RGM is also fractional step, but both steps are
MCMs
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Chorin’s RVM for 2D Navier-Stokes:
P
1. Using stream function “blobs” represent ψ(x) = Nn=1 ξi ψ 0 (x − xi )
2. Advect each of the vortex “blobs” ∆t forward in time via the ODEs
for the x and y components of ξi , use your favorite (stable) ODE
method:
X ∂ψ 0 (xi − xj )
dxi
= −
ξj
, i = 1, . . . , N
dt
∂y
j6=i
dyi
dt
=
X
j6=i
ξj
∂ψ 0 (xi − xj )
, i = 1, . . . , N
∂x
3. Update each xi (t + ∆t) → xi (t + ∆t) + ηi where ηi is N(0, 2∆tγ)
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Chorin’s RVM for 2D Navier-Stokes:
P
1. Using stream function “blobs” represent ψ(x) = Nn=1 ξi ψ 0 (x − xi )
2. Advect each of the vortex “blobs” ∆t forward in time via the ODEs
for the x and y components of ξi , use your favorite (stable) ODE
method:
X ∂ψ 0 (xi − xj )
dxi
= −
ξj
, i = 1, . . . , N
dt
∂y
j6=i
dyi
dt
=
X
j6=i
ξj
∂ψ 0 (xi − xj )
, i = 1, . . . , N
∂x
3. Update each xi (t + ∆t) → xi (t + ∆t) + ηi where ηi is N(0, 2∆tγ)
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Chorin’s RVM for 2D Navier-Stokes:
P
1. Using stream function “blobs” represent ψ(x) = Nn=1 ξi ψ 0 (x − xi )
2. Advect each of the vortex “blobs” ∆t forward in time via the ODEs
for the x and y components of ξi , use your favorite (stable) ODE
method:
X ∂ψ 0 (xi − xj )
dxi
= −
ξj
, i = 1, . . . , N
dt
∂y
j6=i
dyi
dt
=
X
j6=i
ξj
∂ψ 0 (xi − xj )
, i = 1, . . . , N
∂x
3. Update each xi (t + ∆t) → xi (t + ∆t) + ηi where ηi is N(0, 2∆tγ)
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Chorin’s RVM for 2D Navier-Stokes:
P
1. Using stream function “blobs” represent ψ(x) = Nn=1 ξi ψ 0 (x − xi )
2. Advect each of the vortex “blobs” ∆t forward in time via the ODEs
for the x and y components of ξi , use your favorite (stable) ODE
method:
X ∂ψ 0 (xi − xj )
dxi
= −
ξj
, i = 1, . . . , N
dt
∂y
j6=i
dyi
dt
=
X
j6=i
ξj
∂ψ 0 (xi − xj )
, i = 1, . . . , N
∂x
3. Update each xi (t + ∆t) → xi (t + ∆t) + ηi where ηi is N(0, 2∆tγ)
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Shortcoming of method is irrotational flows, i.e. ∇ × u = 0,
i.e. ξ = 0
1. Irrotational flow implies ∃φ s.t. u = ∇φ, i.e. u is a potential flow
2. 2D potential flows reduce to solving ∇ · u = ∆φ = 0 with boundary
conditions
3. Can use this to enforce boundary conditions, as can add a
potential flow that corrects for the rotational flow at the boundaries
I
As with RGM the RVM is adaptive in that regions of high vorticity
(rotation) get a high density of vortex “blobs”
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Shortcoming of method is irrotational flows, i.e. ∇ × u = 0,
i.e. ξ = 0
1. Irrotational flow implies ∃φ s.t. u = ∇φ, i.e. u is a potential flow
2. 2D potential flows reduce to solving ∇ · u = ∆φ = 0 with boundary
conditions
3. Can use this to enforce boundary conditions, as can add a
potential flow that corrects for the rotational flow at the boundaries
I
As with RGM the RVM is adaptive in that regions of high vorticity
(rotation) get a high density of vortex “blobs”
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Shortcoming of method is irrotational flows, i.e. ∇ × u = 0,
i.e. ξ = 0
1. Irrotational flow implies ∃φ s.t. u = ∇φ, i.e. u is a potential flow
2. 2D potential flows reduce to solving ∇ · u = ∆φ = 0 with boundary
conditions
3. Can use this to enforce boundary conditions, as can add a
potential flow that corrects for the rotational flow at the boundaries
I
As with RGM the RVM is adaptive in that regions of high vorticity
(rotation) get a high density of vortex “blobs”
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Shortcoming of method is irrotational flows, i.e. ∇ × u = 0,
i.e. ξ = 0
1. Irrotational flow implies ∃φ s.t. u = ∇φ, i.e. u is a potential flow
2. 2D potential flows reduce to solving ∇ · u = ∆φ = 0 with boundary
conditions
3. Can use this to enforce boundary conditions, as can add a
potential flow that corrects for the rotational flow at the boundaries
I
As with RGM the RVM is adaptive in that regions of high vorticity
(rotation) get a high density of vortex “blobs”
MCM for PDEs
Probabilistic Representations of PDEs
Monte Carlo Methods for PDEs from Fluid Mechanics
Chorin’s RVM for the Incompressible Navier-Stokes
Equations
I
Shortcoming of method is irrotational flows, i.e. ∇ × u = 0,
i.e. ξ = 0
1. Irrotational flow implies ∃φ s.t. u = ∇φ, i.e. u is a potential flow
2. 2D potential flows reduce to solving ∇ · u = ∆φ = 0 with boundary
conditions
3. Can use this to enforce boundary conditions, as can add a
potential flow that corrects for the rotational flow at the boundaries
I
As with RGM the RVM is adaptive in that regions of high vorticity
(rotation) get a high density of vortex “blobs”
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
We have constructed MCMs for both elliptic and parabolic PDEs
but have not considered MCMs for hyperbolic PDEs except for
Berger’s equation (was a very special case)
I
In general MCMs for hyperbolic PDEs (like the wave equation:
utt = c 2 uxx ) are hard to derive as Brownian motion is
fundamentally related to diffusion (parabolic PDEs) and to the
equilibrium of diffusion processes (elliptic PDEs), in contrast
hyperbolic problems model distortion free information
propagation which is fundamentally nonrandom
I
A famous special case of an hyperbolic MCM for the
telegrapher’s equation (Kac, 1956):
2a ∂F
1 ∂2F
+ 2
= ∆F ,
c 2 ∂t 2
c ∂t
∂F (x, 0)
F (x, 0) = φ(x),
=0
∂t
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
We have constructed MCMs for both elliptic and parabolic PDEs
but have not considered MCMs for hyperbolic PDEs except for
Berger’s equation (was a very special case)
I
In general MCMs for hyperbolic PDEs (like the wave equation:
utt = c 2 uxx ) are hard to derive as Brownian motion is
fundamentally related to diffusion (parabolic PDEs) and to the
equilibrium of diffusion processes (elliptic PDEs), in contrast
hyperbolic problems model distortion free information
propagation which is fundamentally nonrandom
I
A famous special case of an hyperbolic MCM for the
telegrapher’s equation (Kac, 1956):
1 ∂2F
2a ∂F
+ 2
= ∆F ,
c 2 ∂t 2
c ∂t
∂F (x, 0)
F (x, 0) = φ(x),
=0
∂t
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
We have constructed MCMs for both elliptic and parabolic PDEs
but have not considered MCMs for hyperbolic PDEs except for
Berger’s equation (was a very special case)
I
In general MCMs for hyperbolic PDEs (like the wave equation:
utt = c 2 uxx ) are hard to derive as Brownian motion is
fundamentally related to diffusion (parabolic PDEs) and to the
equilibrium of diffusion processes (elliptic PDEs), in contrast
hyperbolic problems model distortion free information
propagation which is fundamentally nonrandom
I
A famous special case of an hyperbolic MCM for the
telegrapher’s equation (Kac, 1956):
1 ∂2F
2a ∂F
+ 2
= ∆F ,
c 2 ∂t 2
c ∂t
∂F (x, 0)
F (x, 0) = φ(x),
=0
∂t
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
The telegrapher’s equation approaches both the wave and heat
equations in different limiting cases
1. Wave equation: a → 0
2. Heat equation: a, c → ∞, 2a/c 2 →
1
D
I
Consider the one-dimensional telegrapher’s equation, when
a = 0 we know the solution is given by F (x, t) = φ(x+ct)+φ(x−ct)
2
I
If we think of a as the probability per unit time of a Poisson
process then N(t) = # of events occurring up to time t has the
k
distribution P{N(t) = k } = e−at (at)
k!
Rt
If a particle moves with velocity c for time t it travels ct = 0 c dτ ,
if it undergoes random Poisson distributed direction reversal with
probability per unit time a, the distance traveled in time t is
Rt
c(−1)N(τ ) dτ
0
I
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
The telegrapher’s equation approaches both the wave and heat
equations in different limiting cases
1. Wave equation: a → 0
2. Heat equation: a, c → ∞, 2a/c 2 →
1
D
I
Consider the one-dimensional telegrapher’s equation, when
a = 0 we know the solution is given by F (x, t) = φ(x+ct)+φ(x−ct)
2
I
If we think of a as the probability per unit time of a Poisson
process then N(t) = # of events occurring up to time t has the
k
distribution P{N(t) = k } = e−at (at)
k!
Rt
If a particle moves with velocity c for time t it travels ct = 0 c dτ ,
if it undergoes random Poisson distributed direction reversal with
probability per unit time a, the distance traveled in time t is
Rt
c(−1)N(τ ) dτ
0
I
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
The telegrapher’s equation approaches both the wave and heat
equations in different limiting cases
1. Wave equation: a → 0
2. Heat equation: a, c → ∞, 2a/c 2 →
1
D
I
Consider the one-dimensional telegrapher’s equation, when
a = 0 we know the solution is given by F (x, t) = φ(x+ct)+φ(x−ct)
2
I
If we think of a as the probability per unit time of a Poisson
process then N(t) = # of events occurring up to time t has the
k
distribution P{N(t) = k } = e−at (at)
k!
Rt
If a particle moves with velocity c for time t it travels ct = 0 c dτ ,
if it undergoes random Poisson distributed direction reversal with
probability per unit time a, the distance traveled in time t is
Rt
c(−1)N(τ ) dτ
0
I
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
The telegrapher’s equation approaches both the wave and heat
equations in different limiting cases
1. Wave equation: a → 0
2. Heat equation: a, c → ∞, 2a/c 2 →
1
D
I
Consider the one-dimensional telegrapher’s equation, when
a = 0 we know the solution is given by F (x, t) = φ(x+ct)+φ(x−ct)
2
I
If we think of a as the probability per unit time of a Poisson
process then N(t) = # of events occurring up to time t has the
k
distribution P{N(t) = k } = e−at (at)
k!
Rt
If a particle moves with velocity c for time t it travels ct = 0 c dτ ,
if it undergoes random Poisson distributed direction reversal with
probability per unit time a, the distance traveled in time t is
Rt
c(−1)N(τ ) dτ
0
I
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
The telegrapher’s equation approaches both the wave and heat
equations in different limiting cases
1. Wave equation: a → 0
2. Heat equation: a, c → ∞, 2a/c 2 →
1
D
I
Consider the one-dimensional telegrapher’s equation, when
a = 0 we know the solution is given by F (x, t) = φ(x+ct)+φ(x−ct)
2
I
If we think of a as the probability per unit time of a Poisson
process then N(t) = # of events occurring up to time t has the
k
distribution P{N(t) = k } = e−at (at)
k!
Rt
If a particle moves with velocity c for time t it travels ct = 0 c dτ ,
if it undergoes random Poisson distributed direction reversal with
probability per unit time a, the distance traveled in time t is
Rt
c(−1)N(τ ) dτ
0
I
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
The telegrapher’s equation approaches both the wave and heat
equations in different limiting cases
1. Wave equation: a → 0
2. Heat equation: a, c → ∞, 2a/c 2 →
1
D
I
Consider the one-dimensional telegrapher’s equation, when
a = 0 we know the solution is given by F (x, t) = φ(x+ct)+φ(x−ct)
2
I
If we think of a as the probability per unit time of a Poisson
process then N(t) = # of events occurring up to time t has the
k
distribution P{N(t) = k } = e−at (at)
k!
Rt
If a particle moves with velocity c for time t it travels ct = 0 c dτ ,
if it undergoes random Poisson distributed direction reversal with
probability per unit time a, the distance traveled in time t is
Rt
c(−1)N(τ ) dτ
0
I
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
I
I
If we replace ct in the exact solution to the 1D wave equation by
the randomized distance traveled average over all Poisson
reversing paths we get:
!
Z t
1
N(τ )
c(−1)
dτ
+
F (x, t) = E φ x +
2
0
!
Z t
1
N(τ )
E φ x−
c(−1)
dτ
2
0
which can be proven to solve the above IVP for the telegrapher’s
equation
In any dimension, an exact solution for the wave equation can be
converted into a solution to the telegrapher’s equation by
replacing t in the wave equation ansatz by the randomized time
Rt
(−1)N(τ ) dτ and averaging
0
This is the basis of a MCM for the telegrapher’s equation, one
can also construct MCMs for finite-difference approximations to
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
I
I
If we replace ct in the exact solution to the 1D wave equation by
the randomized distance traveled average over all Poisson
reversing paths we get:
!
Z t
1
N(τ )
c(−1)
dτ
+
F (x, t) = E φ x +
2
0
!
Z t
1
N(τ )
E φ x−
c(−1)
dτ
2
0
which can be proven to solve the above IVP for the telegrapher’s
equation
In any dimension, an exact solution for the wave equation can be
converted into a solution to the telegrapher’s equation by
replacing t in the wave equation ansatz by the randomized time
Rt
(−1)N(τ ) dτ and averaging
0
This is the basis of a MCM for the telegrapher’s equation, one
can also construct MCMs for finite-difference approximations to
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
MCMs for Other PDEs
I
I
I
If we replace ct in the exact solution to the 1D wave equation by
the randomized distance traveled average over all Poisson
reversing paths we get:
!
Z t
1
N(τ )
c(−1)
dτ
+
F (x, t) = E φ x +
2
0
!
Z t
1
N(τ )
E φ x−
c(−1)
dτ
2
0
which can be proven to solve the above IVP for the telegrapher’s
equation
In any dimension, an exact solution for the wave equation can be
converted into a solution to the telegrapher’s equation by
replacing t in the wave equation ansatz by the randomized time
Rt
(−1)N(τ ) dτ and averaging
0
This is the basis of a MCM for the telegrapher’s equation, one
can also construct MCMs for finite-difference approximations to
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
Nonlinear Equations and the Feynman-Kac Formula
I
Recall that in the Feynman-Kac formula the operator L enters in
through the SDE which generates the sample paths over which
expected values are taken
I
If one replaces the sample paths from solutions of linear SDEs
with paths derived from branching processes one can sample
certain nonlinear parabolic PDEs directly (McKean, 1988)
I
Recall that the solution to the IVP for the heat equation is
represented via Feynman-Kac as: u(x, t) = Ex [u0 (β(t))]
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
Nonlinear Equations and the Feynman-Kac Formula
I
Recall that in the Feynman-Kac formula the operator L enters in
through the SDE which generates the sample paths over which
expected values are taken
I
If one replaces the sample paths from solutions of linear SDEs
with paths derived from branching processes one can sample
certain nonlinear parabolic PDEs directly (McKean, 1988)
I
Recall that the solution to the IVP for the heat equation is
represented via Feynman-Kac as: u(x, t) = Ex [u0 (β(t))]
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
Nonlinear Equations and the Feynman-Kac Formula
I
Recall that in the Feynman-Kac formula the operator L enters in
through the SDE which generates the sample paths over which
expected values are taken
I
If one replaces the sample paths from solutions of linear SDEs
with paths derived from branching processes one can sample
certain nonlinear parabolic PDEs directly (McKean, 1988)
I
Recall that the solution to the IVP for the heat equation is
represented via Feynman-Kac as: u(x, t) = Ex [u0 (β(t))]
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
Nonlinear Equations and the Feynman-Kac Formula
Consider normal Brownian motion with exponentially distributed
branching with unit branching probability per unit time, then the
Feynman-Kac representation above with expectations taken over this
branching process instead of normal Brownian motion solves the IVP
for the Kolmogorov-Petrovskii-Piskunov equation:
1
∂u
= ∆u + u 2 − u , u(x, 0) = u0 (x)
∂t
2
has solution
u(x, t) = Ex
N(t)
Y
u0 (xi (t))
i=1
where the branching Brownian motion started at x at t = 0 leads to
N(t) particles at time t with locations xi (t), i = 1, . . . , N(t)
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
Nonlinear Equations and the Feynman-Kac Formula
I
If instead of binary branching at exponentially distributed
P∞
branching times there are n with probability pn , n=2 pn = 1,
then the branching Feynman-Kac formula solves the IVP for:
!
∞
X
∂u
1
n
= ∆u +
pn u − u
∂t
2
n=2
I
P∞
Can also have pn < 0 with n=2 |pn | = 1 with same result except
that must have two representations for the normal and the
“negative” walkers
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
Nonlinear Equations and the Feynman-Kac Formula
I
If instead of binary branching at exponentially distributed
P∞
branching times there are n with probability pn , n=2 pn = 1,
then the branching Feynman-Kac formula solves the IVP for:
!
∞
X
∂u
1
n
= ∆u +
pn u − u
∂t
2
n=2
I
P∞
Can also have pn < 0 with n=2 |pn | = 1 with same result except
that must have two representations for the normal and the
“negative” walkers
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
Nonlinear Equations and the Feynman-Kac Formula
I
If λ = the probability per unit time of branching then we can solve
the IVP for:
!
∞
X
1
∂u
n
= ∆u + λ
pn u − u
∂t
2
n=2
I
If we have k (x, t) as inhomogeneous branching probability per
unit time, then we solve the IVP for:
!
Z t
∞
X
∂u
1
n
= ∆u +
k (x, τ ) dτ
pn u − u
∂t
2
0
n=2
MCM for PDEs
Probabilistic Representations of PDEs
Probabilistic Representations for Other PDEs
Nonlinear Equations and the Feynman-Kac Formula
I
If λ = the probability per unit time of branching then we can solve
the IVP for:
!
∞
X
1
∂u
n
= ∆u + λ
pn u − u
∂t
2
n=2
I
If we have k (x, t) as inhomogeneous branching probability per
unit time, then we solve the IVP for:
!
Z t
∞
X
∂u
1
n
= ∆u +
k (x, τ ) dτ
pn u − u
∂t
2
0
n=2
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Monte Carlo has been, and continues to be used in linear algebra
I
Consider the linear system: x = Hx + b, if ||H|| = H < 1, then
the following iterative method converges:
x n+1 := Hx n + b, x 0 = 0,
Pk −1
and in particular we have x k = i=0 H i b, and similarly the
Neumann series converges:
N=
∞
X
−1
i
H = (I − H)
i=0
I
I
,
||N|| =
∞
X
i
||H || ≤
i=0
Formally, the solution is x = (I − H)−1 b
Note: ||H|| can be defined in many ways, e. g.:
1. ||H|| = maxi
P
j
|hij |
2. ||H|| = maxi (|λi (H)|) = ρ(H) (spectral radius)
∞
X
i=0
Hi =
1
1−H
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Monte Carlo has been, and continues to be used in linear algebra
I
Consider the linear system: x = Hx + b, if ||H|| = H < 1, then
the following iterative method converges:
x n+1 := Hx n + b, x 0 = 0,
Pk −1
and in particular we have x k = i=0 H i b, and similarly the
Neumann series converges:
N=
∞
X
−1
i
H = (I − H)
i=0
I
I
,
||N|| =
∞
X
i
||H || ≤
i=0
Formally, the solution is x = (I − H)−1 b
Note: ||H|| can be defined in many ways, e. g.:
1. ||H|| = maxi
P
j
|hij |
2. ||H|| = maxi (|λi (H)|) = ρ(H) (spectral radius)
∞
X
i=0
Hi =
1
1−H
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Monte Carlo has been, and continues to be used in linear algebra
I
Consider the linear system: x = Hx + b, if ||H|| = H < 1, then
the following iterative method converges:
x n+1 := Hx n + b, x 0 = 0,
Pk −1
and in particular we have x k = i=0 H i b, and similarly the
Neumann series converges:
N=
∞
X
−1
i
H = (I − H)
i=0
I
I
,
||N|| =
∞
X
i
||H || ≤
i=0
Formally, the solution is x = (I − H)−1 b
Note: ||H|| can be defined in many ways, e. g.:
1. ||H|| = maxi
P
j
|hij |
2. ||H|| = maxi (|λi (H)|) = ρ(H) (spectral radius)
∞
X
i=0
Hi =
1
1−H
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Monte Carlo has been, and continues to be used in linear algebra
I
Consider the linear system: x = Hx + b, if ||H|| = H < 1, then
the following iterative method converges:
x n+1 := Hx n + b, x 0 = 0,
Pk −1
and in particular we have x k = i=0 H i b, and similarly the
Neumann series converges:
N=
∞
X
−1
i
H = (I − H)
i=0
I
I
,
||N|| =
∞
X
i
||H || ≤
i=0
Formally, the solution is x = (I − H)−1 b
Note: ||H|| can be defined in many ways, e. g.:
1. ||H|| = maxi
P
j
|hij |
2. ||H|| = maxi (|λi (H)|) = ρ(H) (spectral radius)
∞
X
i=0
Hi =
1
1−H
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Monte Carlo has been, and continues to be used in linear algebra
I
Consider the linear system: x = Hx + b, if ||H|| = H < 1, then
the following iterative method converges:
x n+1 := Hx n + b, x 0 = 0,
Pk −1
and in particular we have x k = i=0 H i b, and similarly the
Neumann series converges:
N=
∞
X
−1
i
H = (I − H)
i=0
I
I
,
||N|| =
∞
X
i
||H || ≤
i=0
Formally, the solution is x = (I − H)−1 b
Note: ||H|| can be defined in many ways, e. g.:
1. ||H|| = maxi
P
j
|hij |
2. ||H|| = maxi (|λi (H)|) = ρ(H) (spectral radius)
∞
X
i=0
Hi =
1
1−H
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Monte Carlo has been, and continues to be used in linear algebra
I
Consider the linear system: x = Hx + b, if ||H|| = H < 1, then
the following iterative method converges:
x n+1 := Hx n + b, x 0 = 0,
Pk −1
and in particular we have x k = i=0 H i b, and similarly the
Neumann series converges:
N=
∞
X
−1
i
H = (I − H)
i=0
I
I
,
||N|| =
∞
X
i
||H || ≤
i=0
Formally, the solution is x = (I − H)−1 b
Note: ||H|| can be defined in many ways, e. g.:
1. ||H|| = maxi
P
j
|hij |
2. ||H|| = maxi (|λi (H)|) = ρ(H) (spectral radius)
∞
X
i=0
Hi =
1
1−H
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Recall: Monte Carlo can be used to form a sum: S =
follows
PM
i=1
ai as
1. P
Define pi ≥ 0 as the probability of choosing index i, with
M
i=1 pi = 1, and pi > 0 whenever ai 6= 0
2. Then ai /pi with index ichosen
with {pi } is an unbiased estimate of
P
ai
p
=
S
S, as E[ai /pi ] = M
i
i=1 pi
PM ai2 3. Var [ai /pi ] ∝ i=1 p , so optimal choice is pi ∝ ai2
i
I
Given ||H|| = H < 1, solving x = Hx + b via Neumann series
requires successive matrix-vector multiplication
I
We can use the above sampling of a sum to construct a sample
of the Neumann series
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Recall: Monte Carlo can be used to form a sum: S =
follows
PM
i=1
ai as
1. P
Define pi ≥ 0 as the probability of choosing index i, with
M
i=1 pi = 1, and pi > 0 whenever ai 6= 0
2. Then ai /pi with index ichosen
with {pi } is an unbiased estimate of
P
ai
p
=
S
S, as E[ai /pi ] = M
i
i=1 pi
PM ai2 3. Var [ai /pi ] ∝ i=1 p , so optimal choice is pi ∝ ai2
i
I
Given ||H|| = H < 1, solving x = Hx + b via Neumann series
requires successive matrix-vector multiplication
I
We can use the above sampling of a sum to construct a sample
of the Neumann series
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Recall: Monte Carlo can be used to form a sum: S =
follows
PM
i=1
ai as
1. P
Define pi ≥ 0 as the probability of choosing index i, with
M
i=1 pi = 1, and pi > 0 whenever ai 6= 0
2. Then ai /pi with index ichosen
with {pi } is an unbiased estimate of
P
ai
p
=
S
S, as E[ai /pi ] = M
i
i=1 pi
PM ai2 3. Var [ai /pi ] ∝ i=1 p , so optimal choice is pi ∝ ai2
i
I
Given ||H|| = H < 1, solving x = Hx + b via Neumann series
requires successive matrix-vector multiplication
I
We can use the above sampling of a sum to construct a sample
of the Neumann series
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Recall: Monte Carlo can be used to form a sum: S =
follows
PM
i=1
ai as
1. P
Define pi ≥ 0 as the probability of choosing index i, with
M
i=1 pi = 1, and pi > 0 whenever ai 6= 0
2. Then ai /pi with index ichosen
with {pi } is an unbiased estimate of
P
ai
p
=
S
S, as E[ai /pi ] = M
i
i=1 pi
PM ai2 3. Var [ai /pi ] ∝ i=1 p , so optimal choice is pi ∝ ai2
i
I
Given ||H|| = H < 1, solving x = Hx + b via Neumann series
requires successive matrix-vector multiplication
I
We can use the above sampling of a sum to construct a sample
of the Neumann series
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Recall: Monte Carlo can be used to form a sum: S =
follows
PM
i=1
ai as
1. P
Define pi ≥ 0 as the probability of choosing index i, with
M
i=1 pi = 1, and pi > 0 whenever ai 6= 0
2. Then ai /pi with index ichosen
with {pi } is an unbiased estimate of
P
ai
p
=
S
S, as E[ai /pi ] = M
i
i=1 pi
PM ai2 3. Var [ai /pi ] ∝ i=1 p , so optimal choice is pi ∝ ai2
i
I
Given ||H|| = H < 1, solving x = Hx + b via Neumann series
requires successive matrix-vector multiplication
I
We can use the above sampling of a sum to construct a sample
of the Neumann series
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Recall: Monte Carlo can be used to form a sum: S =
follows
PM
i=1
ai as
1. P
Define pi ≥ 0 as the probability of choosing index i, with
M
i=1 pi = 1, and pi > 0 whenever ai 6= 0
2. Then ai /pi with index ichosen
with {pi } is an unbiased estimate of
P
ai
p
=
S
S, as E[ai /pi ] = M
i
i=1 pi
PM ai2 3. Var [ai /pi ] ∝ i=1 p , so optimal choice is pi ∝ ai2
i
I
Given ||H|| = H < 1, solving x = Hx + b via Neumann series
requires successive matrix-vector multiplication
I
We can use the above sampling of a sum to construct a sample
of the Neumann series
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
The Ulam-von Neumann Method
I
I
We first construct a Markov chain based on H and b to sample a
`-fold product matrices as follows
Define the transition probability matrix, P
1. pij ≥ 0, and pij >P
0 when hij 6= 0
2. Define pi = 1 − j pij
I
Now define a Markov chain on states {0, 1, . . . , n} with transition
probability, pij , and termination probability from state i, pi
I
Also define
(h
ij
vij =
pij
0,
,
hij 6= 0
otherwise
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
The Ulam-von Neumann Method
I
I
We first construct a Markov chain based on H and b to sample a
`-fold product matrices as follows
Define the transition probability matrix, P
1. pij ≥ 0, and pij >P
0 when hij 6= 0
2. Define pi = 1 − j pij
I
Now define a Markov chain on states {0, 1, . . . , n} with transition
probability, pij , and termination probability from state i, pi
I
Also define
(h
ij
vij =
pij
0,
,
hij 6= 0
otherwise
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
The Ulam-von Neumann Method
I
I
We first construct a Markov chain based on H and b to sample a
`-fold product matrices as follows
Define the transition probability matrix, P
1. pij ≥ 0, and pij >P
0 when hij 6= 0
2. Define pi = 1 − j pij
I
Now define a Markov chain on states {0, 1, . . . , n} with transition
probability, pij , and termination probability from state i, pi
I
Also define
(h
ij
vij =
pij
0,
,
hij 6= 0
otherwise
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
The Ulam-von Neumann Method
I
I
We first construct a Markov chain based on H and b to sample a
`-fold product matrices as follows
Define the transition probability matrix, P
1. pij ≥ 0, and pij >P
0 when hij 6= 0
2. Define pi = 1 − j pij
I
Now define a Markov chain on states {0, 1, . . . , n} with transition
probability, pij , and termination probability from state i, pi
I
Also define
(h
ij
vij =
pij
0,
,
hij 6= 0
otherwise
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
The Ulam-von Neumann Method
I
I
We first construct a Markov chain based on H and b to sample a
`-fold product matrices as follows
Define the transition probability matrix, P
1. pij ≥ 0, and pij >P
0 when hij 6= 0
2. Define pi = 1 − j pij
I
Now define a Markov chain on states {0, 1, . . . , n} with transition
probability, pij , and termination probability from state i, pi
I
Also define
(h
ij
vij =
pij
0,
,
hij 6= 0
otherwise
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
The Ulam-von Neumann Method
I
I
We first construct a Markov chain based on H and b to sample a
`-fold product matrices as follows
Define the transition probability matrix, P
1. pij ≥ 0, and pij >P
0 when hij 6= 0
2. Define pi = 1 − j pij
I
Now define a Markov chain on states {0, 1, . . . , n} with transition
probability, pij , and termination probability from state i, pi
I
Also define
(h
ij
vij =
pij
0,
,
hij 6= 0
otherwise
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Given the desire to sample xi we create the following estimator
based on
1. Generating a Markov chain γ(i0 , i1 , . . . , ik ) using pij and pi , where
state ik is the penultimate before absorbtion
2. Form the estimator
m
X(γ) = Vm (γ)
X
bik
, where Vm (γ) =
vij−1 ij , m ≤ k
pik
j=1
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Given the desire to sample xi we create the following estimator
based on
1. Generating a Markov chain γ(i0 , i1 , . . . , ik ) using pij and pi , where
state ik is the penultimate before absorbtion
2. Form the estimator
m
X(γ) = Vm (γ)
X
bik
, where Vm (γ) =
vij−1 ij , m ≤ k
pik
j=1
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
Given the desire to sample xi we create the following estimator
based on
1. Generating a Markov chain γ(i0 , i1 , . . . , ik ) using pij and pi , where
state ik is the penultimate before absorbtion
2. Form the estimator
m
X(γ) = Vm (γ)
X
bik
, where Vm (γ) =
vij−1 ij , m ≤ k
pik
j=1
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
I
We have
E [X(γ)|i0 = i] =
∞
X
k =0
X

k =0
∞
X

i1
···
X
ik

bik 
pii1 . . . pik −1 ik vii1 . . . vik −1 ik
=
pik


X X
b
i

···
pii1 vii1 . . . pik −1 ik vik −1 ik k  =
pik
i1
ik


∞
∞
X
X X
X

···
hii1 hik −1 ik bik  = the ith component of
Hk b
k =0
i1
ik
k =0
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
Elaborations on Monte Carlo for Linear Algebra
I
Backward walks (Adjoint sampling)
I
The lower variance Wasow estimator
X∗ (γ) =
k
X
Vm (γ)bik
m=0
I
I
I
Again with only matrix-vector multiplication can do the power
method and shifted variants for sampling eigenvalues
Newer, more highly convergent methods are based on randomly
sampling smaller problems
All these have integral equation and Feynman-Kac analogs
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
Elaborations on Monte Carlo for Linear Algebra
I
Backward walks (Adjoint sampling)
I
The lower variance Wasow estimator
X∗ (γ) =
k
X
Vm (γ)bik
m=0
I
I
I
Again with only matrix-vector multiplication can do the power
method and shifted variants for sampling eigenvalues
Newer, more highly convergent methods are based on randomly
sampling smaller problems
All these have integral equation and Feynman-Kac analogs
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
Elaborations on Monte Carlo for Linear Algebra
I
Backward walks (Adjoint sampling)
I
The lower variance Wasow estimator
X∗ (γ) =
k
X
Vm (γ)bik
m=0
I
I
I
Again with only matrix-vector multiplication can do the power
method and shifted variants for sampling eigenvalues
Newer, more highly convergent methods are based on randomly
sampling smaller problems
All these have integral equation and Feynman-Kac analogs
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
Elaborations on Monte Carlo for Linear Algebra
I
Backward walks (Adjoint sampling)
I
The lower variance Wasow estimator
X∗ (γ) =
k
X
Vm (γ)bik
m=0
I
I
I
Again with only matrix-vector multiplication can do the power
method and shifted variants for sampling eigenvalues
Newer, more highly convergent methods are based on randomly
sampling smaller problems
All these have integral equation and Feynman-Kac analogs
MCM for PDEs
Monte Carlo Methods and Linear Algebra
Monte Carlo Methods and Linear Algebra
Elaborations on Monte Carlo for Linear Algebra
I
Backward walks (Adjoint sampling)
I
The lower variance Wasow estimator
X∗ (γ) =
k
X
Vm (γ)bik
m=0
I
I
I
Again with only matrix-vector multiplication can do the power
method and shifted variants for sampling eigenvalues
Newer, more highly convergent methods are based on randomly
sampling smaller problems
All these have integral equation and Feynman-Kac analogs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
I
I
Now that we know all these MCMs, we must discuss how to
implement these methods on parallel (and vector) computers
There are two major classes of parallel computer being
commercially produced
1. Single Instruction Multiple Data machines:
I
Only one data stream so the same instruction is broadcast to all
processors
I
Usually these machines have many simple processors, often
they are bit serial (fine grained)
I
Usually these machines have distributed memory
I
Connection Machine and vector-processing units are “classical”
examples
I
Modern examples include GPGPUs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
I
I
Now that we know all these MCMs, we must discuss how to
implement these methods on parallel (and vector) computers
There are two major classes of parallel computer being
commercially produced
1. Single Instruction Multiple Data machines:
I
Only one data stream so the same instruction is broadcast to all
processors
I
Usually these machines have many simple processors, often
they are bit serial (fine grained)
I
Usually these machines have distributed memory
I
Connection Machine and vector-processing units are “classical”
examples
I
Modern examples include GPGPUs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
I
I
Now that we know all these MCMs, we must discuss how to
implement these methods on parallel (and vector) computers
There are two major classes of parallel computer being
commercially produced
1. Single Instruction Multiple Data machines:
I
Only one data stream so the same instruction is broadcast to all
processors
I
Usually these machines have many simple processors, often
they are bit serial (fine grained)
I
Usually these machines have distributed memory
I
Connection Machine and vector-processing units are “classical”
examples
I
Modern examples include GPGPUs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
I
I
Now that we know all these MCMs, we must discuss how to
implement these methods on parallel (and vector) computers
There are two major classes of parallel computer being
commercially produced
1. Single Instruction Multiple Data machines:
I
Only one data stream so the same instruction is broadcast to all
processors
I
Usually these machines have many simple processors, often
they are bit serial (fine grained)
I
Usually these machines have distributed memory
I
Connection Machine and vector-processing units are “classical”
examples
I
Modern examples include GPGPUs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
I
I
Now that we know all these MCMs, we must discuss how to
implement these methods on parallel (and vector) computers
There are two major classes of parallel computer being
commercially produced
1. Single Instruction Multiple Data machines:
I
Only one data stream so the same instruction is broadcast to all
processors
I
Usually these machines have many simple processors, often
they are bit serial (fine grained)
I
Usually these machines have distributed memory
I
Connection Machine and vector-processing units are “classical”
examples
I
Modern examples include GPGPUs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
I
I
Now that we know all these MCMs, we must discuss how to
implement these methods on parallel (and vector) computers
There are two major classes of parallel computer being
commercially produced
1. Single Instruction Multiple Data machines:
I
Only one data stream so the same instruction is broadcast to all
processors
I
Usually these machines have many simple processors, often
they are bit serial (fine grained)
I
Usually these machines have distributed memory
I
Connection Machine and vector-processing units are “classical”
examples
I
Modern examples include GPGPUs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
I
I
Now that we know all these MCMs, we must discuss how to
implement these methods on parallel (and vector) computers
There are two major classes of parallel computer being
commercially produced
1. Single Instruction Multiple Data machines:
I
Only one data stream so the same instruction is broadcast to all
processors
I
Usually these machines have many simple processors, often
they are bit serial (fine grained)
I
Usually these machines have distributed memory
I
Connection Machine and vector-processing units are “classical”
examples
I
Modern examples include GPGPUs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
I
I
Now that we know all these MCMs, we must discuss how to
implement these methods on parallel (and vector) computers
There are two major classes of parallel computer being
commercially produced
1. Single Instruction Multiple Data machines:
I
Only one data stream so the same instruction is broadcast to all
processors
I
Usually these machines have many simple processors, often
they are bit serial (fine grained)
I
Usually these machines have distributed memory
I
Connection Machine and vector-processing units are “classical”
examples
I
Modern examples include GPGPUs
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
Parallel Computing Overview
1. Multiple Instruction Multiple Data machines:
I
I
I
I
I
I
I
These machines are a collection of conventional computers with an
interconnection network
Each processor can run its own program
Processors are usually large grained
Can have shared or distributed memory
Shared memory limits the number of processors though bus
technology
Distributed memory can be implemented with many different
interconnection topologies
Modern examples:
1.1 Multicore machines
1.2 Clustered machines
1.3 Shared-memory machines
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Principles for Constructing Parallel Algorithms
I
Use widely replicated aspects of a given problem as the basis for
parallelism
1. In MCMs can map each independent statistical sample to an
independent process with essentially no interprocessor
communication
2. With spatial systems, spatial subdomains (domain decomposition)
can be used but implicit schemes and global conditions (elliptic
BVPs) will require considerable communication
3. In some cases solutions over time are desired as in ODE problems,
and the iterative solution of the entire time trajectory (continuation
methods) can be distributed by time point
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Principles for Constructing Parallel Algorithms
I
Use widely replicated aspects of a given problem as the basis for
parallelism
1. In MCMs can map each independent statistical sample to an
independent process with essentially no interprocessor
communication
2. With spatial systems, spatial subdomains (domain decomposition)
can be used but implicit schemes and global conditions (elliptic
BVPs) will require considerable communication
3. In some cases solutions over time are desired as in ODE problems,
and the iterative solution of the entire time trajectory (continuation
methods) can be distributed by time point
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Principles for Constructing Parallel Algorithms
I
Use widely replicated aspects of a given problem as the basis for
parallelism
1. In MCMs can map each independent statistical sample to an
independent process with essentially no interprocessor
communication
2. With spatial systems, spatial subdomains (domain decomposition)
can be used but implicit schemes and global conditions (elliptic
BVPs) will require considerable communication
3. In some cases solutions over time are desired as in ODE problems,
and the iterative solution of the entire time trajectory (continuation
methods) can be distributed by time point
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Principles for Constructing Parallel Algorithms
I
Use widely replicated aspects of a given problem as the basis for
parallelism
1. In MCMs can map each independent statistical sample to an
independent process with essentially no interprocessor
communication
2. With spatial systems, spatial subdomains (domain decomposition)
can be used but implicit schemes and global conditions (elliptic
BVPs) will require considerable communication
3. In some cases solutions over time are desired as in ODE problems,
and the iterative solution of the entire time trajectory (continuation
methods) can be distributed by time point
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Approaches to the Construction of MCMs for
Elliptic BVPs
I
As a simple example, consider the Dirichlet problem for the
Laplace equation (1)
I
The Wiener integral representation for the solution to (1) is
u(x) = Ex [g(β(τ∂Ω ))]
I
General approaches are to use (i) “exact” samples of the Wiener
integral, (ii) sample from a discretization of the Wiener integral
I
Can sample with spherical processes, use the MVP and spheres
to walk until from the boundary, other elliptic L’s lead to
uniformity on ellipses instead of spheres (i)
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Approaches to the Construction of MCMs for
Elliptic BVPs
I
As a simple example, consider the Dirichlet problem for the
Laplace equation (1)
I
The Wiener integral representation for the solution to (1) is
u(x) = Ex [g(β(τ∂Ω ))]
I
General approaches are to use (i) “exact” samples of the Wiener
integral, (ii) sample from a discretization of the Wiener integral
I
Can sample with spherical processes, use the MVP and spheres
to walk until from the boundary, other elliptic L’s lead to
uniformity on ellipses instead of spheres (i)
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Approaches to the Construction of MCMs for
Elliptic BVPs
I
As a simple example, consider the Dirichlet problem for the
Laplace equation (1)
I
The Wiener integral representation for the solution to (1) is
u(x) = Ex [g(β(τ∂Ω ))]
I
General approaches are to use (i) “exact” samples of the Wiener
integral, (ii) sample from a discretization of the Wiener integral
I
Can sample with spherical processes, use the MVP and spheres
to walk until from the boundary, other elliptic L’s lead to
uniformity on ellipses instead of spheres (i)
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Approaches to the Construction of MCMs for
Elliptic BVPs
I
As a simple example, consider the Dirichlet problem for the
Laplace equation (1)
I
The Wiener integral representation for the solution to (1) is
u(x) = Ex [g(β(τ∂Ω ))]
I
General approaches are to use (i) “exact” samples of the Wiener
integral, (ii) sample from a discretization of the Wiener integral
I
Can sample with spherical processes, use the MVP and spheres
to walk until from the boundary, other elliptic L’s lead to
uniformity on ellipses instead of spheres (i)
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Approaches to the Construction of MCMs for
Elliptic BVPs
I
Random Fourier series relationship to Brownian motion can be
used to generate walks via a truncated series, in some cases this
gives an exact random series solution which is then sampled,
with complicated Ω this approach is not practical (ii)
I
Can use a high dimensional integral approximation of the
infinite-dimensional (Wiener) integral, the finite-dimensional
integral is then evaluated via analytic or MCMs (ii)
I
Spatially discretize the region and sample from the discrete
Wiener integral (ii)
I
This last is a very fruitful approach especially w.r.t. parallel
computers
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Approaches to the Construction of MCMs for
Elliptic BVPs
I
Random Fourier series relationship to Brownian motion can be
used to generate walks via a truncated series, in some cases this
gives an exact random series solution which is then sampled,
with complicated Ω this approach is not practical (ii)
I
Can use a high dimensional integral approximation of the
infinite-dimensional (Wiener) integral, the finite-dimensional
integral is then evaluated via analytic or MCMs (ii)
I
Spatially discretize the region and sample from the discrete
Wiener integral (ii)
I
This last is a very fruitful approach especially w.r.t. parallel
computers
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Approaches to the Construction of MCMs for
Elliptic BVPs
I
Random Fourier series relationship to Brownian motion can be
used to generate walks via a truncated series, in some cases this
gives an exact random series solution which is then sampled,
with complicated Ω this approach is not practical (ii)
I
Can use a high dimensional integral approximation of the
infinite-dimensional (Wiener) integral, the finite-dimensional
integral is then evaluated via analytic or MCMs (ii)
I
Spatially discretize the region and sample from the discrete
Wiener integral (ii)
I
This last is a very fruitful approach especially w.r.t. parallel
computers
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
General Approaches to the Construction of MCMs for
Elliptic BVPs
I
Random Fourier series relationship to Brownian motion can be
used to generate walks via a truncated series, in some cases this
gives an exact random series solution which is then sampled,
with complicated Ω this approach is not practical (ii)
I
Can use a high dimensional integral approximation of the
infinite-dimensional (Wiener) integral, the finite-dimensional
integral is then evaluated via analytic or MCMs (ii)
I
Spatially discretize the region and sample from the discrete
Wiener integral (ii)
I
This last is a very fruitful approach especially w.r.t. parallel
computers
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Discrete Wiener Integrals
I
All of the theory for continuous sample path Wiener integrals
mentioned above carries over to the discrete cases
1. Can replace the continuous region Ω with a discretization Ωh where
h the characteristic discretization parameter
2. Replace β(·) with random walks on Ωh , βh (·), this requires a
transition probability matrix for the walks on the grid [P]ij = pij
3. E.g. the discrete Wiener integral solution to equation (1):
uh (x) = Exh [g(βh (τ∂Ωh ))]
4. In this case if one has elliptic regularity of u(x) and a nonsingular
discretization, Ωh then uh (x) = u(x) + O(h2 )
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Discrete Wiener Integrals
I
All of the theory for continuous sample path Wiener integrals
mentioned above carries over to the discrete cases
1. Can replace the continuous region Ω with a discretization Ωh where
h the characteristic discretization parameter
2. Replace β(·) with random walks on Ωh , βh (·), this requires a
transition probability matrix for the walks on the grid [P]ij = pij
3. E.g. the discrete Wiener integral solution to equation (1):
uh (x) = Exh [g(βh (τ∂Ωh ))]
4. In this case if one has elliptic regularity of u(x) and a nonsingular
discretization, Ωh then uh (x) = u(x) + O(h2 )
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Discrete Wiener Integrals
I
All of the theory for continuous sample path Wiener integrals
mentioned above carries over to the discrete cases
1. Can replace the continuous region Ω with a discretization Ωh where
h the characteristic discretization parameter
2. Replace β(·) with random walks on Ωh , βh (·), this requires a
transition probability matrix for the walks on the grid [P]ij = pij
3. E.g. the discrete Wiener integral solution to equation (1):
uh (x) = Exh [g(βh (τ∂Ωh ))]
4. In this case if one has elliptic regularity of u(x) and a nonsingular
discretization, Ωh then uh (x) = u(x) + O(h2 )
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Discrete Wiener Integrals
I
All of the theory for continuous sample path Wiener integrals
mentioned above carries over to the discrete cases
1. Can replace the continuous region Ω with a discretization Ωh where
h the characteristic discretization parameter
2. Replace β(·) with random walks on Ωh , βh (·), this requires a
transition probability matrix for the walks on the grid [P]ij = pij
3. E.g. the discrete Wiener integral solution to equation (1):
uh (x) = Exh [g(βh (τ∂Ωh ))]
4. In this case if one has elliptic regularity of u(x) and a nonsingular
discretization, Ωh then uh (x) = u(x) + O(h2 )
MCM for PDEs
Parallel Computing Overview
General Principles for Constructing Parallel Algorithms
Discrete Wiener Integrals
I
All of the theory for continuous sample path Wiener integrals
mentioned above carries over to the discrete cases
1. Can replace the continuous region Ω with a discretization Ωh where
h the characteristic discretization parameter
2. Replace β(·) with random walks on Ωh , βh (·), this requires a
transition probability matrix for the walks on the grid [P]ij = pij
3. E.g. the discrete Wiener integral solution to equation (1):
uh (x) = Exh [g(βh (τ∂Ωh ))]
4. In this case if one has elliptic regularity of u(x) and a nonsingular
discretization, Ωh then uh (x) = u(x) + O(h2 )
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel N-body Potential Evaluation
I
N-body potential problems are common in biochemistry, stellar
dynamics, fluid dynamics, materials simulation, the boundary
element method for elliptic PDEs, etc.
I
The solution of N-body potential problems requires evaluation of
PN
function of the form: Φ(x) = n=1 φ(x − xi ) for all values of
x = xj , j = 1, . . . , N
I
One heuristic solution is to replace φ(x) with a cutoff version
φco (x) = φ(x) for |x| ≤ r and φco (x) = 0 otherwise, this reduces
the problem to only keeping track of r -neighborhood points
I
Can use the xi , xj interaction as the basis for parallelism and use
N 2 processors to calculate the N(N − 1)/2 terms in parallel,
initialization of the coordinates and accumulation the results
requires O(log2 N) operations
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel N-body Potential Evaluation
I
N-body potential problems are common in biochemistry, stellar
dynamics, fluid dynamics, materials simulation, the boundary
element method for elliptic PDEs, etc.
I
The solution of N-body potential problems requires evaluation of
PN
function of the form: Φ(x) = n=1 φ(x − xi ) for all values of
x = xj , j = 1, . . . , N
I
One heuristic solution is to replace φ(x) with a cutoff version
φco (x) = φ(x) for |x| ≤ r and φco (x) = 0 otherwise, this reduces
the problem to only keeping track of r -neighborhood points
I
Can use the xi , xj interaction as the basis for parallelism and use
N 2 processors to calculate the N(N − 1)/2 terms in parallel,
initialization of the coordinates and accumulation the results
requires O(log2 N) operations
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel N-body Potential Evaluation
I
N-body potential problems are common in biochemistry, stellar
dynamics, fluid dynamics, materials simulation, the boundary
element method for elliptic PDEs, etc.
I
The solution of N-body potential problems requires evaluation of
PN
function of the form: Φ(x) = n=1 φ(x − xi ) for all values of
x = xj , j = 1, . . . , N
I
One heuristic solution is to replace φ(x) with a cutoff version
φco (x) = φ(x) for |x| ≤ r and φco (x) = 0 otherwise, this reduces
the problem to only keeping track of r -neighborhood points
I
Can use the xi , xj interaction as the basis for parallelism and use
N 2 processors to calculate the N(N − 1)/2 terms in parallel,
initialization of the coordinates and accumulation the results
requires O(log2 N) operations
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel N-body Potential Evaluation
I
N-body potential problems are common in biochemistry, stellar
dynamics, fluid dynamics, materials simulation, the boundary
element method for elliptic PDEs, etc.
I
The solution of N-body potential problems requires evaluation of
PN
function of the form: Φ(x) = n=1 φ(x − xi ) for all values of
x = xj , j = 1, . . . , N
I
One heuristic solution is to replace φ(x) with a cutoff version
φco (x) = φ(x) for |x| ≤ r and φco (x) = 0 otherwise, this reduces
the problem to only keeping track of r -neighborhood points
I
Can use the xi , xj interaction as the basis for parallelism and use
N 2 processors to calculate the N(N − 1)/2 terms in parallel,
initialization of the coordinates and accumulation the results
requires O(log2 N) operations
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel N-body Potential Evaluation
I
The fast multipole method is an efficiency improvement over
direct methods
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
I
Algorithm is based on multipole expansion and some theory from
complex variable series, consider the electrostatic description
If z ∈ C then a charge of intensity q at z0 results in a complex
potential via Laurent series for |z| > |z0 |:
∞
X
1 z0 k
φz0 (z) = q ln(z − z0 ) = q ln(z) −
k z
k =1
1. If we have m charges qi at locations zi then the potential induced
by them is given by the multipole expansion for
|z| > r = maxi |zi |:
∞
X
ak
,
φ(z) = Q ln(z) +
zk
k =1
where Q
=
m
X
i=1
qi , and ak =
m
X
−qi z k
i
i=1
k
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
I
Algorithm is based on multipole expansion and some theory from
complex variable series, consider the electrostatic description
If z ∈ C then a charge of intensity q at z0 results in a complex
potential via Laurent series for |z| > |z0 |:
∞
X
1 z0 k
φz0 (z) = q ln(z − z0 ) = q ln(z) −
k z
k =1
1. If we have m charges qi at locations zi then the potential induced
by them is given by the multipole expansion for
|z| > r = maxi |zi |:
∞
X
ak
φ(z) = Q ln(z) +
,
zk
k =1
where Q
=
m
X
i=1
qi , and ak =
m
X
−qi z k
i
i=1
k
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
I
Algorithm is based on multipole expansion and some theory from
complex variable series, consider the electrostatic description
If z ∈ C then a charge of intensity q at z0 results in a complex
potential via Laurent series for |z| > |z0 |:
∞
X
1 z0 k
φz0 (z) = q ln(z − z0 ) = q ln(z) −
k z
k =1
1. If we have m charges qi at locations zi then the potential induced
by them is given by the multipole expansion for
|z| > r = maxi |zi |:
∞
X
ak
φ(z) = Q ln(z) +
,
zk
k =1
where Q
=
m
X
i=1
qi , and ak =
m
X
−qi z k
i
i=1
k
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Given an accuracy, , one can truncate the multipole expansion
to a fixed number, p = d− logc e, of terms, where c = | zr |
I
With p determined one can store a multipole expansion as
{a1 , a2 . . . , ap }
1. We can move a multipole’s center from z0 to the origin with new
coefficients:
l
X
Qz l
l −1
bl = − 0 +
ak z0l−k
l
k −1
k =1
I
Note that {a1 , a2 . . . , ap } can be used to exactly compute
{b1 , b2 . . . , bp }
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Given an accuracy, , one can truncate the multipole expansion
to a fixed number, p = d− logc e, of terms, where c = | zr |
I
With p determined one can store a multipole expansion as
{a1 , a2 . . . , ap }
1. We can move a multipole’s center from z0 to the origin with new
coefficients:
l
X
Qz l
l −1
bl = − 0 +
ak z0l−k
l
k −1
k =1
I
Note that {a1 , a2 . . . , ap } can be used to exactly compute
{b1 , b2 . . . , bp }
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Given an accuracy, , one can truncate the multipole expansion
to a fixed number, p = d− logc e, of terms, where c = | zr |
I
With p determined one can store a multipole expansion as
{a1 , a2 . . . , ap }
1. We can move a multipole’s center from z0 to the origin with new
coefficients:
l
X
Qz l
l −1
bl = − 0 +
ak z0l−k
l
k −1
k =1
I
Note that {a1 , a2 . . . , ap } can be used to exactly compute
{b1 , b2 . . . , bp }
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Given an accuracy, , one can truncate the multipole expansion
to a fixed number, p = d− logc e, of terms, where c = | zr |
I
With p determined one can store a multipole expansion as
{a1 , a2 . . . , ap }
1. We can move a multipole’s center from z0 to the origin with new
coefficients:
l
X
Qz l
l −1
bl = − 0 +
ak z0l−k
l
k −1
k =1
I
Note that {a1 , a2 . . . , ap } can be used to exactly compute
{b1 , b2 . . . , bp }
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
1. Can also convert a multipole Laurent expansion into a local
Taylor expansion:
∞
X
φ(z) =
cl z l , where
l=0
c0
=
cl
=
∞
X
ak
(−1)k , and
k
z
k =0 0
∞
X
1
ak l + k − 1
Q
− l + l
(−1)k
k −1
l ż0
z0 k =1 z0k
Q ln(−z0 ) +
2. And translate local expansions:
!
n
n
n
X
X
X
k
k −l
k
ak
(−z0 )
zl
ak (z − z0 ) =
l
k =0
l=0
k =l
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
1. Can also convert a multipole Laurent expansion into a local
Taylor expansion:
∞
X
φ(z) =
cl z l , where
l=0
c0
=
cl
=
∞
X
ak
(−1)k , and
k
z
k =0 0
∞
X
1
ak l + k − 1
Q
− l + l
(−1)k
k −1
l ż0
z0 k =1 z0k
Q ln(−z0 ) +
2. And translate local expansions:
!
n
n
n
X
X
X
k
k −l
k
ak
(−z0 )
zl
ak (z − z0 ) =
l
k =0
l=0
k =l
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
I
I
Items (1)-(4) above are the machinery required to allow the
construction and use of multipole expansions, and given a
multipole expansion, it requires O(N) operations to evaluate it at
N points, thus an algorithm for the construction of a multipole
expansion from N point charges that requires O(N) operations
reduces the complexity of N-body problems to O(N) complexity
The Rokhlin-Greengard algorithm achieves this by using a
multiscale approach and (1)-(4)
Consider a box enclosing z1 , z2 , . . . , zN , and n ≈ dlog4 Ne
refinements of the box, in 2D one parent box becomes four
children boxes
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
I
I
Items (1)-(4) above are the machinery required to allow the
construction and use of multipole expansions, and given a
multipole expansion, it requires O(N) operations to evaluate it at
N points, thus an algorithm for the construction of a multipole
expansion from N point charges that requires O(N) operations
reduces the complexity of N-body problems to O(N) complexity
The Rokhlin-Greengard algorithm achieves this by using a
multiscale approach and (1)-(4)
Consider a box enclosing z1 , z2 , . . . , zN , and n ≈ dlog4 Ne
refinements of the box, in 2D one parent box becomes four
children boxes
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
I
I
Items (1)-(4) above are the machinery required to allow the
construction and use of multipole expansions, and given a
multipole expansion, it requires O(N) operations to evaluate it at
N points, thus an algorithm for the construction of a multipole
expansion from N point charges that requires O(N) operations
reduces the complexity of N-body problems to O(N) complexity
The Rokhlin-Greengard algorithm achieves this by using a
multiscale approach and (1)-(4)
Consider a box enclosing z1 , z2 , . . . , zN , and n ≈ dlog4 Ne
refinements of the box, in 2D one parent box becomes four
children boxes
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Goal is to construct all p-term multipole expansions due to the
particles in each box and level (upward pass) and then use these
to construct local expansions in each box and level (downward
pass)
1. The upward pass:
I
I
At the finest level construct box-centered p-term multipole expansions
due to the particles in each box using (1)
At each coarser level shift child p-term multipole expansions to build
box-centered p-term multipole expansions due to the particles in the
parent boxes using (2)
2. The downward pass:
I
Construct local Taylor expansion in each box at the finest level by
converting the p-term multipole expansions of boxes in the “interaction
list” via (3)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Goal is to construct all p-term multipole expansions due to the
particles in each box and level (upward pass) and then use these
to construct local expansions in each box and level (downward
pass)
1. The upward pass:
I
I
At the finest level construct box-centered p-term multipole expansions
due to the particles in each box using (1)
At each coarser level shift child p-term multipole expansions to build
box-centered p-term multipole expansions due to the particles in the
parent boxes using (2)
2. The downward pass:
I
Construct local Taylor expansion in each box at the finest level by
converting the p-term multipole expansions of boxes in the “interaction
list” via (3)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Goal is to construct all p-term multipole expansions due to the
particles in each box and level (upward pass) and then use these
to construct local expansions in each box and level (downward
pass)
1. The upward pass:
I
I
At the finest level construct box-centered p-term multipole expansions
due to the particles in each box using (1)
At each coarser level shift child p-term multipole expansions to build
box-centered p-term multipole expansions due to the particles in the
parent boxes using (2)
2. The downward pass:
I
Construct local Taylor expansion in each box at the finest level by
converting the p-term multipole expansions of boxes in the “interaction
list” via (3)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Goal is to construct all p-term multipole expansions due to the
particles in each box and level (upward pass) and then use these
to construct local expansions in each box and level (downward
pass)
1. The upward pass:
I
I
At the finest level construct box-centered p-term multipole expansions
due to the particles in each box using (1)
At each coarser level shift child p-term multipole expansions to build
box-centered p-term multipole expansions due to the particles in the
parent boxes using (2)
2. The downward pass:
I
Construct local Taylor expansion in each box at the finest level by
converting the p-term multipole expansions of boxes in the “interaction
list” via (3)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Goal is to construct all p-term multipole expansions due to the
particles in each box and level (upward pass) and then use these
to construct local expansions in each box and level (downward
pass)
1. The upward pass:
I
I
At the finest level construct box-centered p-term multipole expansions
due to the particles in each box using (1)
At each coarser level shift child p-term multipole expansions to build
box-centered p-term multipole expansions due to the particles in the
parent boxes using (2)
2. The downward pass:
I
Construct local Taylor expansion in each box at the finest level by
converting the p-term multipole expansions of boxes in the “interaction
list” via (3)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Goal is to construct all p-term multipole expansions due to the
particles in each box and level (upward pass) and then use these
to construct local expansions in each box and level (downward
pass)
1. The upward pass:
I
I
At the finest level construct box-centered p-term multipole expansions
due to the particles in each box using (1)
At each coarser level shift child p-term multipole expansions to build
box-centered p-term multipole expansions due to the particles in the
parent boxes using (2)
2. The downward pass:
I
Construct local Taylor expansion in each box at the finest level by
converting the p-term multipole expansions of boxes in the “interaction
list” via (3)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
1. The downward pass (cont.):
I
I
I
Via (4) add these together to get local box-centered expansions for
all the particles outside the box’s neighborhood using coarser level
p-term multipole expansions
Evaluate the above local expansions at each particle and add in
the directly compute nearest-neighbor interactions
This algorithm is overall O(N) complex and has many parallel
aspects
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
1. The downward pass (cont.):
I
I
I
Via (4) add these together to get local box-centered expansions for
all the particles outside the box’s neighborhood using coarser level
p-term multipole expansions
Evaluate the above local expansions at each particle and add in
the directly compute nearest-neighbor interactions
This algorithm is overall O(N) complex and has many parallel
aspects
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
1. The downward pass (cont.):
I
I
I
Via (4) add these together to get local box-centered expansions for
all the particles outside the box’s neighborhood using coarser level
p-term multipole expansions
Evaluate the above local expansions at each particle and add in
the directly compute nearest-neighbor interactions
This algorithm is overall O(N) complex and has many parallel
aspects
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
1. The downward pass (cont.):
I
I
I
Via (4) add these together to get local box-centered expansions for
all the particles outside the box’s neighborhood using coarser level
p-term multipole expansions
Evaluate the above local expansions at each particle and add in
the directly compute nearest-neighbor interactions
This algorithm is overall O(N) complex and has many parallel
aspects
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Need only store the p-term multipole/local Taylor expansion
coefficients {a1 , a2 , . . . , ap }, giving a trivial data structure with
which to implementationally deal
I
This version of the multipole algorithm depends on rather uniform
distribution of particles, however there is an adaptive version
where the “finest boxes” are allowed to be different sizes
s.t. each box has approximately one particle
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
The Rokhlin-Greengard Fast Multipole Method
I
Need only store the p-term multipole/local Taylor expansion
coefficients {a1 , a2 , . . . , ap }, giving a trivial data structure with
which to implementationally deal
I
This version of the multipole algorithm depends on rather uniform
distribution of particles, however there is an adaptive version
where the “finest boxes” are allowed to be different sizes
s.t. each box has approximately one particle
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel Implementation of the Fast Multipole Method
I
In all the steps described in both the upward and downward
passes of the multipole method the multipole or local expansions
can all be done in parallel
I
Thus the parallel complexity is proportional to the number of
levels, i.e. O(log4 N) = O(n)
I
For the nonadaptive version there a serious load balancing
problem due to whether at the each level, l, each box contains
exactly N/nl particles
I
This is a multigrid algorithm and so there will be the problem of
idle processors at coarse levels on parallel machines with fine
grainsize
I
One can also implement the adaptive version of the multipole
method in parallel
I
For more detail on the parallel version see (Greengard and
Gropp, 1990)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel Implementation of the Fast Multipole Method
I
In all the steps described in both the upward and downward
passes of the multipole method the multipole or local expansions
can all be done in parallel
I
Thus the parallel complexity is proportional to the number of
levels, i.e. O(log4 N) = O(n)
I
For the nonadaptive version there a serious load balancing
problem due to whether at the each level, l, each box contains
exactly N/nl particles
I
This is a multigrid algorithm and so there will be the problem of
idle processors at coarse levels on parallel machines with fine
grainsize
I
One can also implement the adaptive version of the multipole
method in parallel
I
For more detail on the parallel version see (Greengard and
Gropp, 1990)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel Implementation of the Fast Multipole Method
I
In all the steps described in both the upward and downward
passes of the multipole method the multipole or local expansions
can all be done in parallel
I
Thus the parallel complexity is proportional to the number of
levels, i.e. O(log4 N) = O(n)
I
For the nonadaptive version there a serious load balancing
problem due to whether at the each level, l, each box contains
exactly N/nl particles
I
This is a multigrid algorithm and so there will be the problem of
idle processors at coarse levels on parallel machines with fine
grainsize
I
One can also implement the adaptive version of the multipole
method in parallel
I
For more detail on the parallel version see (Greengard and
Gropp, 1990)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel Implementation of the Fast Multipole Method
I
In all the steps described in both the upward and downward
passes of the multipole method the multipole or local expansions
can all be done in parallel
I
Thus the parallel complexity is proportional to the number of
levels, i.e. O(log4 N) = O(n)
I
For the nonadaptive version there a serious load balancing
problem due to whether at the each level, l, each box contains
exactly N/nl particles
I
This is a multigrid algorithm and so there will be the problem of
idle processors at coarse levels on parallel machines with fine
grainsize
I
One can also implement the adaptive version of the multipole
method in parallel
I
For more detail on the parallel version see (Greengard and
Gropp, 1990)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel Implementation of the Fast Multipole Method
I
In all the steps described in both the upward and downward
passes of the multipole method the multipole or local expansions
can all be done in parallel
I
Thus the parallel complexity is proportional to the number of
levels, i.e. O(log4 N) = O(n)
I
For the nonadaptive version there a serious load balancing
problem due to whether at the each level, l, each box contains
exactly N/nl particles
I
This is a multigrid algorithm and so there will be the problem of
idle processors at coarse levels on parallel machines with fine
grainsize
I
One can also implement the adaptive version of the multipole
method in parallel
I
For more detail on the parallel version see (Greengard and
Gropp, 1990)
MCM for PDEs
Parallel Computing Overview
Parallel N-body Potential Evaluation
Parallel Implementation of the Fast Multipole Method
I
In all the steps described in both the upward and downward
passes of the multipole method the multipole or local expansions
can all be done in parallel
I
Thus the parallel complexity is proportional to the number of
levels, i.e. O(log4 N) = O(n)
I
For the nonadaptive version there a serious load balancing
problem due to whether at the each level, l, each box contains
exactly N/nl particles
I
This is a multigrid algorithm and so there will be the problem of
idle processors at coarse levels on parallel machines with fine
grainsize
I
One can also implement the adaptive version of the multipole
method in parallel
I
For more detail on the parallel version see (Greengard and
Gropp, 1990)
MCM for PDEs
Bibliography
Bibliography
Booth, T. E. (1981) “Exact Monte Carlo solutions of elliptic partial
differential equations," J. Comp. Phys., 39: 396–404.
Brandt A. (1977) “Multi-level adaptive solutions to boundary value
problems," Math. Comp., 31: 333–390.
Chorin, A. J. (1973) "Numerical study of slightly viscous flow,"
J. Fluid Mech., 57: 785–796.
Chorin, Alexandre J. and Hald, Ole H. (2006) Stochastic Tools in
Mathematics and Science, Surveys and Tutorials in the Applied
Mathematical Sciences, Vol. 1, viii+147, Springer, New York.
Courant, R., K. O. Friedrichs, and H. Lewy (1928) “Über die
partiellen Differenzengleichungen der mathematischen Physik,"
Math. Ann., 100: 32–74 (In German), Reprinted in
I.B.M. J. Res. and Dev., 11: 215–234 (In English).
Curtiss, J. H. (1953) “Monte Carlo methods for the iteration of
linear operators," J. Math. and Phys., 32: 209–232.
MCM for PDEs
Bibliography
Bibliography
Booth, T. E. (1981) “Exact Monte Carlo solutions of elliptic partial
differential equations," J. Comp. Phys., 39: 396–404.
Brandt A. (1977) “Multi-level adaptive solutions to boundary value
problems," Math. Comp., 31: 333–390.
Chorin, A. J. (1973) "Numerical study of slightly viscous flow,"
J. Fluid Mech., 57: 785–796.
Chorin, Alexandre J. and Hald, Ole H. (2006) Stochastic Tools in
Mathematics and Science, Surveys and Tutorials in the Applied
Mathematical Sciences, Vol. 1, viii+147, Springer, New York.
Courant, R., K. O. Friedrichs, and H. Lewy (1928) “Über die
partiellen Differenzengleichungen der mathematischen Physik,"
Math. Ann., 100: 32–74 (In German), Reprinted in
I.B.M. J. Res. and Dev., 11: 215–234 (In English).
Curtiss, J. H. (1953) “Monte Carlo methods for the iteration of
linear operators," J. Math. and Phys., 32: 209–232.
MCM for PDEs
Bibliography
Bibliography
Booth, T. E. (1981) “Exact Monte Carlo solutions of elliptic partial
differential equations," J. Comp. Phys., 39: 396–404.
Brandt A. (1977) “Multi-level adaptive solutions to boundary value
problems," Math. Comp., 31: 333–390.
Chorin, A. J. (1973) "Numerical study of slightly viscous flow,"
J. Fluid Mech., 57: 785–796.
Chorin, Alexandre J. and Hald, Ole H. (2006) Stochastic Tools in
Mathematics and Science, Surveys and Tutorials in the Applied
Mathematical Sciences, Vol. 1, viii+147, Springer, New York.
Courant, R., K. O. Friedrichs, and H. Lewy (1928) “Über die
partiellen Differenzengleichungen der mathematischen Physik,"
Math. Ann., 100: 32–74 (In German), Reprinted in
I.B.M. J. Res. and Dev., 11: 215–234 (In English).
Curtiss, J. H. (1953) “Monte Carlo methods for the iteration of
linear operators," J. Math. and Phys., 32: 209–232.
MCM for PDEs
Bibliography
Bibliography
Booth, T. E. (1981) “Exact Monte Carlo solutions of elliptic partial
differential equations," J. Comp. Phys., 39: 396–404.
Brandt A. (1977) “Multi-level adaptive solutions to boundary value
problems," Math. Comp., 31: 333–390.
Chorin, A. J. (1973) "Numerical study of slightly viscous flow,"
J. Fluid Mech., 57: 785–796.
Chorin, Alexandre J. and Hald, Ole H. (2006) Stochastic Tools in
Mathematics and Science, Surveys and Tutorials in the Applied
Mathematical Sciences, Vol. 1, viii+147, Springer, New York.
Courant, R., K. O. Friedrichs, and H. Lewy (1928) “Über die
partiellen Differenzengleichungen der mathematischen Physik,"
Math. Ann., 100: 32–74 (In German), Reprinted in
I.B.M. J. Res. and Dev., 11: 215–234 (In English).
Curtiss, J. H. (1953) “Monte Carlo methods for the iteration of
linear operators," J. Math. and Phys., 32: 209–232.
MCM for PDEs
Bibliography
Bibliography
Booth, T. E. (1981) “Exact Monte Carlo solutions of elliptic partial
differential equations," J. Comp. Phys., 39: 396–404.
Brandt A. (1977) “Multi-level adaptive solutions to boundary value
problems," Math. Comp., 31: 333–390.
Chorin, A. J. (1973) "Numerical study of slightly viscous flow,"
J. Fluid Mech., 57: 785–796.
Chorin, Alexandre J. and Hald, Ole H. (2006) Stochastic Tools in
Mathematics and Science, Surveys and Tutorials in the Applied
Mathematical Sciences, Vol. 1, viii+147, Springer, New York.
Courant, R., K. O. Friedrichs, and H. Lewy (1928) “Über die
partiellen Differenzengleichungen der mathematischen Physik,"
Math. Ann., 100: 32–74 (In German), Reprinted in
I.B.M. J. Res. and Dev., 11: 215–234 (In English).
Curtiss, J. H. (1953) “Monte Carlo methods for the iteration of
linear operators," J. Math. and Phys., 32: 209–232.
MCM for PDEs
Bibliography
Bibliography
Booth, T. E. (1981) “Exact Monte Carlo solutions of elliptic partial
differential equations," J. Comp. Phys., 39: 396–404.
Brandt A. (1977) “Multi-level adaptive solutions to boundary value
problems," Math. Comp., 31: 333–390.
Chorin, A. J. (1973) "Numerical study of slightly viscous flow,"
J. Fluid Mech., 57: 785–796.
Chorin, Alexandre J. and Hald, Ole H. (2006) Stochastic Tools in
Mathematics and Science, Surveys and Tutorials in the Applied
Mathematical Sciences, Vol. 1, viii+147, Springer, New York.
Courant, R., K. O. Friedrichs, and H. Lewy (1928) “Über die
partiellen Differenzengleichungen der mathematischen Physik,"
Math. Ann., 100: 32–74 (In German), Reprinted in
I.B.M. J. Res. and Dev., 11: 215–234 (In English).
Curtiss, J. H. (1953) “Monte Carlo methods for the iteration of
linear operators," J. Math. and Phys., 32: 209–232.
MCM for PDEs
Bibliography
Bibliography
Curtiss, J. H. (1956) “A theoretical comparison of the efficiencies
of two classical methods and a Monte Carlo method for
computing one component of the solution of a set of linear
algebraic equations," in Symp. Monte Carlo Methods,
H. A. Meyer, Ed., Wiley: New York, pp. 191–233.
Donsker, M. D., and M. Kac (1951) “A sampling method for
determining the lowest eigenvalue and the principle
eigenfunction of Schrödinger’s equation,"
J. Res. Nat. Bur. Standards, 44: 551–557.
Ermakov, S. M. and G. A. Mikhailov (1982) Statistical Modeling,
Nauka, Moscow, (in Russian).
Ermakov, S. M., V. V. Nekrutkin and A. S. Sipin (1989) Random
Processes for Classical Equations of Mathematical Physics,
Kluwer Academic Publishers: Dordrecht.
Forsythe, G. E., and R. A. Leibler (1950) “Matrix inversion by a
Monte Carlo method," Math. Tab. Aids Comput., 4: 127–129.
MCM for PDEs
Bibliography
Bibliography
Curtiss, J. H. (1956) “A theoretical comparison of the efficiencies
of two classical methods and a Monte Carlo method for
computing one component of the solution of a set of linear
algebraic equations," in Symp. Monte Carlo Methods,
H. A. Meyer, Ed., Wiley: New York, pp. 191–233.
Donsker, M. D., and M. Kac (1951) “A sampling method for
determining the lowest eigenvalue and the principle
eigenfunction of Schrödinger’s equation,"
J. Res. Nat. Bur. Standards, 44: 551–557.
Ermakov, S. M. and G. A. Mikhailov (1982) Statistical Modeling,
Nauka, Moscow, (in Russian).
Ermakov, S. M., V. V. Nekrutkin and A. S. Sipin (1989) Random
Processes for Classical Equations of Mathematical Physics,
Kluwer Academic Publishers: Dordrecht.
Forsythe, G. E., and R. A. Leibler (1950) “Matrix inversion by a
Monte Carlo method," Math. Tab. Aids Comput., 4: 127–129.
MCM for PDEs
Bibliography
Bibliography
Curtiss, J. H. (1956) “A theoretical comparison of the efficiencies
of two classical methods and a Monte Carlo method for
computing one component of the solution of a set of linear
algebraic equations," in Symp. Monte Carlo Methods,
H. A. Meyer, Ed., Wiley: New York, pp. 191–233.
Donsker, M. D., and M. Kac (1951) “A sampling method for
determining the lowest eigenvalue and the principle
eigenfunction of Schrödinger’s equation,"
J. Res. Nat. Bur. Standards, 44: 551–557.
Ermakov, S. M. and G. A. Mikhailov (1982) Statistical Modeling,
Nauka, Moscow, (in Russian).
Ermakov, S. M., V. V. Nekrutkin and A. S. Sipin (1989) Random
Processes for Classical Equations of Mathematical Physics,
Kluwer Academic Publishers: Dordrecht.
Forsythe, G. E., and R. A. Leibler (1950) “Matrix inversion by a
Monte Carlo method," Math. Tab. Aids Comput., 4: 127–129.
MCM for PDEs
Bibliography
Bibliography
Curtiss, J. H. (1956) “A theoretical comparison of the efficiencies
of two classical methods and a Monte Carlo method for
computing one component of the solution of a set of linear
algebraic equations," in Symp. Monte Carlo Methods,
H. A. Meyer, Ed., Wiley: New York, pp. 191–233.
Donsker, M. D., and M. Kac (1951) “A sampling method for
determining the lowest eigenvalue and the principle
eigenfunction of Schrödinger’s equation,"
J. Res. Nat. Bur. Standards, 44: 551–557.
Ermakov, S. M. and G. A. Mikhailov (1982) Statistical Modeling,
Nauka, Moscow, (in Russian).
Ermakov, S. M., V. V. Nekrutkin and A. S. Sipin (1989) Random
Processes for Classical Equations of Mathematical Physics,
Kluwer Academic Publishers: Dordrecht.
Forsythe, G. E., and R. A. Leibler (1950) “Matrix inversion by a
Monte Carlo method," Math. Tab. Aids Comput., 4: 127–129.
MCM for PDEs
Bibliography
Bibliography
Curtiss, J. H. (1956) “A theoretical comparison of the efficiencies
of two classical methods and a Monte Carlo method for
computing one component of the solution of a set of linear
algebraic equations," in Symp. Monte Carlo Methods,
H. A. Meyer, Ed., Wiley: New York, pp. 191–233.
Donsker, M. D., and M. Kac (1951) “A sampling method for
determining the lowest eigenvalue and the principle
eigenfunction of Schrödinger’s equation,"
J. Res. Nat. Bur. Standards, 44: 551–557.
Ermakov, S. M. and G. A. Mikhailov (1982) Statistical Modeling,
Nauka, Moscow, (in Russian).
Ermakov, S. M., V. V. Nekrutkin and A. S. Sipin (1989) Random
Processes for Classical Equations of Mathematical Physics,
Kluwer Academic Publishers: Dordrecht.
Forsythe, G. E., and R. A. Leibler (1950) “Matrix inversion by a
Monte Carlo method," Math. Tab. Aids Comput., 4: 127–129.
MCM for PDEs
Bibliography
Bibliography
Freidlin, M. (1985) Functional Integration and Partial Differential
Equations, Princeton University Press: Princeton.
Ghoniem, A. F. and F. S. Sherman (1985) “Grid-free simulation of
diffusion using random walk methods," J. Comp. Phys., 61: 1-37.
Greengard, L. F. (1988) The Rapid Evaluation of Potential Fields
in Particle Systems, MIT Press: Cambridge, MA.
Greengard, L. F. and W. Gropp (1990) “A parallel version of the
fast multipole method," Computers Math. Applic., 20: 63-71.
Hall, A. (1873) “On an experimental determination of π,"
Messeng. Math., 2: 113–114.
Halton, J. H. (1970), “A Retrospective and Prospective Survey of
the Monte Carlo Method," SIAM Review, 12(1): 1–63,.
Hammersley, J. M., and D. C. Handscomb (1964) Monte Carlo
Methods, Chapman and Hall, London.
MCM for PDEs
Bibliography
Bibliography
Freidlin, M. (1985) Functional Integration and Partial Differential
Equations, Princeton University Press: Princeton.
Ghoniem, A. F. and F. S. Sherman (1985) “Grid-free simulation of
diffusion using random walk methods," J. Comp. Phys., 61: 1-37.
Greengard, L. F. (1988) The Rapid Evaluation of Potential Fields
in Particle Systems, MIT Press: Cambridge, MA.
Greengard, L. F. and W. Gropp (1990) “A parallel version of the
fast multipole method," Computers Math. Applic., 20: 63-71.
Hall, A. (1873) “On an experimental determination of π,"
Messeng. Math., 2: 113–114.
Halton, J. H. (1970), “A Retrospective and Prospective Survey of
the Monte Carlo Method," SIAM Review, 12(1): 1–63,.
Hammersley, J. M., and D. C. Handscomb (1964) Monte Carlo
Methods, Chapman and Hall, London.
MCM for PDEs
Bibliography
Bibliography
Freidlin, M. (1985) Functional Integration and Partial Differential
Equations, Princeton University Press: Princeton.
Ghoniem, A. F. and F. S. Sherman (1985) “Grid-free simulation of
diffusion using random walk methods," J. Comp. Phys., 61: 1-37.
Greengard, L. F. (1988) The Rapid Evaluation of Potential Fields
in Particle Systems, MIT Press: Cambridge, MA.
Greengard, L. F. and W. Gropp (1990) “A parallel version of the
fast multipole method," Computers Math. Applic., 20: 63-71.
Hall, A. (1873) “On an experimental determination of π,"
Messeng. Math., 2: 113–114.
Halton, J. H. (1970), “A Retrospective and Prospective Survey of
the Monte Carlo Method," SIAM Review, 12(1): 1–63,.
Hammersley, J. M., and D. C. Handscomb (1964) Monte Carlo
Methods, Chapman and Hall, London.
MCM for PDEs
Bibliography
Bibliography
Freidlin, M. (1985) Functional Integration and Partial Differential
Equations, Princeton University Press: Princeton.
Ghoniem, A. F. and F. S. Sherman (1985) “Grid-free simulation of
diffusion using random walk methods," J. Comp. Phys., 61: 1-37.
Greengard, L. F. (1988) The Rapid Evaluation of Potential Fields
in Particle Systems, MIT Press: Cambridge, MA.
Greengard, L. F. and W. Gropp (1990) “A parallel version of the
fast multipole method," Computers Math. Applic., 20: 63-71.
Hall, A. (1873) “On an experimental determination of π,"
Messeng. Math., 2: 113–114.
Halton, J. H. (1970), “A Retrospective and Prospective Survey of
the Monte Carlo Method," SIAM Review, 12(1): 1–63,.
Hammersley, J. M., and D. C. Handscomb (1964) Monte Carlo
Methods, Chapman and Hall, London.
MCM for PDEs
Bibliography
Bibliography
Freidlin, M. (1985) Functional Integration and Partial Differential
Equations, Princeton University Press: Princeton.
Ghoniem, A. F. and F. S. Sherman (1985) “Grid-free simulation of
diffusion using random walk methods," J. Comp. Phys., 61: 1-37.
Greengard, L. F. (1988) The Rapid Evaluation of Potential Fields
in Particle Systems, MIT Press: Cambridge, MA.
Greengard, L. F. and W. Gropp (1990) “A parallel version of the
fast multipole method," Computers Math. Applic., 20: 63-71.
Hall, A. (1873) “On an experimental determination of π,"
Messeng. Math., 2: 113–114.
Halton, J. H. (1970), “A Retrospective and Prospective Survey of
the Monte Carlo Method," SIAM Review, 12(1): 1–63,.
Hammersley, J. M., and D. C. Handscomb (1964) Monte Carlo
Methods, Chapman and Hall, London.
MCM for PDEs
Bibliography
Bibliography
Freidlin, M. (1985) Functional Integration and Partial Differential
Equations, Princeton University Press: Princeton.
Ghoniem, A. F. and F. S. Sherman (1985) “Grid-free simulation of
diffusion using random walk methods," J. Comp. Phys., 61: 1-37.
Greengard, L. F. (1988) The Rapid Evaluation of Potential Fields
in Particle Systems, MIT Press: Cambridge, MA.
Greengard, L. F. and W. Gropp (1990) “A parallel version of the
fast multipole method," Computers Math. Applic., 20: 63-71.
Hall, A. (1873) “On an experimental determination of π,"
Messeng. Math., 2: 113–114.
Halton, J. H. (1970), “A Retrospective and Prospective Survey of
the Monte Carlo Method," SIAM Review, 12(1): 1–63,.
Hammersley, J. M., and D. C. Handscomb (1964) Monte Carlo
Methods, Chapman and Hall, London.
MCM for PDEs
Bibliography
Bibliography
Freidlin, M. (1985) Functional Integration and Partial Differential
Equations, Princeton University Press: Princeton.
Ghoniem, A. F. and F. S. Sherman (1985) “Grid-free simulation of
diffusion using random walk methods," J. Comp. Phys., 61: 1-37.
Greengard, L. F. (1988) The Rapid Evaluation of Potential Fields
in Particle Systems, MIT Press: Cambridge, MA.
Greengard, L. F. and W. Gropp (1990) “A parallel version of the
fast multipole method," Computers Math. Applic., 20: 63-71.
Hall, A. (1873) “On an experimental determination of π,"
Messeng. Math., 2: 113–114.
Halton, J. H. (1970), “A Retrospective and Prospective Survey of
the Monte Carlo Method," SIAM Review, 12(1): 1–63,.
Hammersley, J. M., and D. C. Handscomb (1964) Monte Carlo
Methods, Chapman and Hall, London.
MCM for PDEs
Bibliography
Bibliography
Halton, J. H. (1989) “Pseudo-random trees: multiple independent
sequence generators for parallel and branching computations,"
J. Comp. Phys., 84: 1–56.
Hillis, D. (1985) The Connection Machine, M.I.T. University Press:
Cambridge, MA.
Hopf, E. (1950) “The partial differential equation ut + uux = µxx ,"
Comm. Pure Applied Math., 3: 201–230.
Itô, K. and H. P. McKean, Jr. (1965) Diffusion Processes and
Their Sample Paths, Springer-Verlag: Berlin, New York.
Kac, M. (1947) “Random Walk and the Theory of Brownian
Motion,"The American Mathematical Monthly, 54(7): 369–391.
Kac, M. (1956) Some Stochastic Problems in Physics and
Mathematics, Colloquium Lectures in the Pure and Applied
Sciences, No. 2, Magnolia Petroleum Co., Hectographed.
Kac, M. (1980) Integration in Function Spaces and Some of Its
Applications, Lezioni Fermiane, Accademia Nazionale Dei Lincei
Scuola Normale Superiore, Pisa.
MCM for PDEs
Bibliography
Bibliography
Halton, J. H. (1989) “Pseudo-random trees: multiple independent
sequence generators for parallel and branching computations,"
J. Comp. Phys., 84: 1–56.
Hillis, D. (1985) The Connection Machine, M.I.T. University Press:
Cambridge, MA.
Hopf, E. (1950) “The partial differential equation ut + uux = µxx ,"
Comm. Pure Applied Math., 3: 201–230.
Itô, K. and H. P. McKean, Jr. (1965) Diffusion Processes and
Their Sample Paths, Springer-Verlag: Berlin, New York.
Kac, M. (1947) “Random Walk and the Theory of Brownian
Motion,"The American Mathematical Monthly, 54(7): 369–391.
Kac, M. (1956) Some Stochastic Problems in Physics and
Mathematics, Colloquium Lectures in the Pure and Applied
Sciences, No. 2, Magnolia Petroleum Co., Hectographed.
Kac, M. (1980) Integration in Function Spaces and Some of Its
Applications, Lezioni Fermiane, Accademia Nazionale Dei Lincei
Scuola Normale Superiore, Pisa.
MCM for PDEs
Bibliography
Bibliography
Halton, J. H. (1989) “Pseudo-random trees: multiple independent
sequence generators for parallel and branching computations,"
J. Comp. Phys., 84: 1–56.
Hillis, D. (1985) The Connection Machine, M.I.T. University Press:
Cambridge, MA.
Hopf, E. (1950) “The partial differential equation ut + uux = µxx ,"
Comm. Pure Applied Math., 3: 201–230.
Itô, K. and H. P. McKean, Jr. (1965) Diffusion Processes and
Their Sample Paths, Springer-Verlag: Berlin, New York.
Kac, M. (1947) “Random Walk and the Theory of Brownian
Motion,"The American Mathematical Monthly, 54(7): 369–391.
Kac, M. (1956) Some Stochastic Problems in Physics and
Mathematics, Colloquium Lectures in the Pure and Applied
Sciences, No. 2, Magnolia Petroleum Co., Hectographed.
Kac, M. (1980) Integration in Function Spaces and Some of Its
Applications, Lezioni Fermiane, Accademia Nazionale Dei Lincei
Scuola Normale Superiore, Pisa.
MCM for PDEs
Bibliography
Bibliography
Halton, J. H. (1989) “Pseudo-random trees: multiple independent
sequence generators for parallel and branching computations,"
J. Comp. Phys., 84: 1–56.
Hillis, D. (1985) The Connection Machine, M.I.T. University Press:
Cambridge, MA.
Hopf, E. (1950) “The partial differential equation ut + uux = µxx ,"
Comm. Pure Applied Math., 3: 201–230.
Itô, K. and H. P. McKean, Jr. (1965) Diffusion Processes and
Their Sample Paths, Springer-Verlag: Berlin, New York.
Kac, M. (1947) “Random Walk and the Theory of Brownian
Motion,"The American Mathematical Monthly, 54(7): 369–391.
Kac, M. (1956) Some Stochastic Problems in Physics and
Mathematics, Colloquium Lectures in the Pure and Applied
Sciences, No. 2, Magnolia Petroleum Co., Hectographed.
Kac, M. (1980) Integration in Function Spaces and Some of Its
Applications, Lezioni Fermiane, Accademia Nazionale Dei Lincei
Scuola Normale Superiore, Pisa.
MCM for PDEs
Bibliography
Bibliography
Halton, J. H. (1989) “Pseudo-random trees: multiple independent
sequence generators for parallel and branching computations,"
J. Comp. Phys., 84: 1–56.
Hillis, D. (1985) The Connection Machine, M.I.T. University Press:
Cambridge, MA.
Hopf, E. (1950) “The partial differential equation ut + uux = µxx ,"
Comm. Pure Applied Math., 3: 201–230.
Itô, K. and H. P. McKean, Jr. (1965) Diffusion Processes and
Their Sample Paths, Springer-Verlag: Berlin, New York.
Kac, M. (1947) “Random Walk and the Theory of Brownian
Motion,"The American Mathematical Monthly, 54(7): 369–391.
Kac, M. (1956) Some Stochastic Problems in Physics and
Mathematics, Colloquium Lectures in the Pure and Applied
Sciences, No. 2, Magnolia Petroleum Co., Hectographed.
Kac, M. (1980) Integration in Function Spaces and Some of Its
Applications, Lezioni Fermiane, Accademia Nazionale Dei Lincei
Scuola Normale Superiore, Pisa.
MCM for PDEs
Bibliography
Bibliography
Halton, J. H. (1989) “Pseudo-random trees: multiple independent
sequence generators for parallel and branching computations,"
J. Comp. Phys., 84: 1–56.
Hillis, D. (1985) The Connection Machine, M.I.T. University Press:
Cambridge, MA.
Hopf, E. (1950) “The partial differential equation ut + uux = µxx ,"
Comm. Pure Applied Math., 3: 201–230.
Itô, K. and H. P. McKean, Jr. (1965) Diffusion Processes and
Their Sample Paths, Springer-Verlag: Berlin, New York.
Kac, M. (1947) “Random Walk and the Theory of Brownian
Motion,"The American Mathematical Monthly, 54(7): 369–391.
Kac, M. (1956) Some Stochastic Problems in Physics and
Mathematics, Colloquium Lectures in the Pure and Applied
Sciences, No. 2, Magnolia Petroleum Co., Hectographed.
Kac, M. (1980) Integration in Function Spaces and Some of Its
Applications, Lezioni Fermiane, Accademia Nazionale Dei Lincei
Scuola Normale Superiore, Pisa.
MCM for PDEs
Bibliography
Bibliography
Halton, J. H. (1989) “Pseudo-random trees: multiple independent
sequence generators for parallel and branching computations,"
J. Comp. Phys., 84: 1–56.
Hillis, D. (1985) The Connection Machine, M.I.T. University Press:
Cambridge, MA.
Hopf, E. (1950) “The partial differential equation ut + uux = µxx ,"
Comm. Pure Applied Math., 3: 201–230.
Itô, K. and H. P. McKean, Jr. (1965) Diffusion Processes and
Their Sample Paths, Springer-Verlag: Berlin, New York.
Kac, M. (1947) “Random Walk and the Theory of Brownian
Motion,"The American Mathematical Monthly, 54(7): 369–391.
Kac, M. (1956) Some Stochastic Problems in Physics and
Mathematics, Colloquium Lectures in the Pure and Applied
Sciences, No. 2, Magnolia Petroleum Co., Hectographed.
Kac, M. (1980) Integration in Function Spaces and Some of Its
Applications, Lezioni Fermiane, Accademia Nazionale Dei Lincei
Scuola Normale Superiore, Pisa.
MCM for PDEs
Bibliography
Bibliography
Knuth, D. E. (1981) The Art of Computer Programming, Vol. 2:
Seminumerical Algorithms, Second edition, Addison-Wesley:
Reading, MA.
Marsaglia, G. and A. Zaman “A new class of random number
generators," submitted to SIAM J. Sci. Stat. Comput.
Mascagni, M. (1991) “High dimensional numerical integration and
massively parallel computing," Contemp. Math., , 115: 53–73.
Mascagni, M. (1990) “A tale of two architectures: parallel Wiener
integral methods for elliptic boundary value problems," SIAM
News, 23: 8,12.
McKean, H. P. (1975 )“Application of Brownian Motion to the
Equation of Kolmogorov-Petrovskii-Piskunov" Communications
on Pure and Applied Mathematics, XXVIII: 323–331.
MCM for PDEs
Bibliography
Bibliography
Knuth, D. E. (1981) The Art of Computer Programming, Vol. 2:
Seminumerical Algorithms, Second edition, Addison-Wesley:
Reading, MA.
Marsaglia, G. and A. Zaman “A new class of random number
generators," submitted to SIAM J. Sci. Stat. Comput.
Mascagni, M. (1991) “High dimensional numerical integration and
massively parallel computing," Contemp. Math., , 115: 53–73.
Mascagni, M. (1990) “A tale of two architectures: parallel Wiener
integral methods for elliptic boundary value problems," SIAM
News, 23: 8,12.
McKean, H. P. (1975 )“Application of Brownian Motion to the
Equation of Kolmogorov-Petrovskii-Piskunov" Communications
on Pure and Applied Mathematics, XXVIII: 323–331.
MCM for PDEs
Bibliography
Bibliography
Knuth, D. E. (1981) The Art of Computer Programming, Vol. 2:
Seminumerical Algorithms, Second edition, Addison-Wesley:
Reading, MA.
Marsaglia, G. and A. Zaman “A new class of random number
generators," submitted to SIAM J. Sci. Stat. Comput.
Mascagni, M. (1991) “High dimensional numerical integration and
massively parallel computing," Contemp. Math., , 115: 53–73.
Mascagni, M. (1990) “A tale of two architectures: parallel Wiener
integral methods for elliptic boundary value problems," SIAM
News, 23: 8,12.
McKean, H. P. (1975 )“Application of Brownian Motion to the
Equation of Kolmogorov-Petrovskii-Piskunov" Communications
on Pure and Applied Mathematics, XXVIII: 323–331.
MCM for PDEs
Bibliography
Bibliography
Knuth, D. E. (1981) The Art of Computer Programming, Vol. 2:
Seminumerical Algorithms, Second edition, Addison-Wesley:
Reading, MA.
Marsaglia, G. and A. Zaman “A new class of random number
generators," submitted to SIAM J. Sci. Stat. Comput.
Mascagni, M. (1991) “High dimensional numerical integration and
massively parallel computing," Contemp. Math., , 115: 53–73.
Mascagni, M. (1990) “A tale of two architectures: parallel Wiener
integral methods for elliptic boundary value problems," SIAM
News, 23: 8,12.
McKean, H. P. (1975 )“Application of Brownian Motion to the
Equation of Kolmogorov-Petrovskii-Piskunov" Communications
on Pure and Applied Mathematics, XXVIII: 323–331.
MCM for PDEs
Bibliography
Bibliography
Knuth, D. E. (1981) The Art of Computer Programming, Vol. 2:
Seminumerical Algorithms, Second edition, Addison-Wesley:
Reading, MA.
Marsaglia, G. and A. Zaman “A new class of random number
generators," submitted to SIAM J. Sci. Stat. Comput.
Mascagni, M. (1991) “High dimensional numerical integration and
massively parallel computing," Contemp. Math., , 115: 53–73.
Mascagni, M. (1990) “A tale of two architectures: parallel Wiener
integral methods for elliptic boundary value problems," SIAM
News, 23: 8,12.
McKean, H. P. (1975 )“Application of Brownian Motion to the
Equation of Kolmogorov-Petrovskii-Piskunov" Communications
on Pure and Applied Mathematics, XXVIII: 323–331.
MCM for PDEs
Bibliography
Bibliography
McKean, H. P. (1988) Private communication. M. E. Muller,
"Some Continuous Monte Carlo Methods for the Dirichlet
Problem," The Annals of Mathematical Statistics, 27(3): 569-589,
1956.
Mikhailov, G. A. (1995) New Monte Carlo Methods With
Estimating Derivatives, V. S. P. Publishers.
Niederreiter, H. (1978) “Quasi-Monte Carlo methods and
pseudo-random numbers," Bull. Amer. Math. Soc., 84: 957–1041.
Rubenstein, M. (1981) Simulation and the Monte Carlo Method,
Wiley-Interscience: New York.
Sabelfeld, K. K. (1991), Monte Carlo Methods in Boundary Value
Problems, Springer-Verlag, Berlin, Heidelberg, New York.
Sabelfeld, K. and N. Mozartova (2009) “Sparsified
Randomization Algorithms for large systems of linear equations
and a new version of the Random Walk on Boundary method,"
Monte Carlo Methods and Applications, 15(3): 257–284.
MCM for PDEs
Bibliography
Bibliography
McKean, H. P. (1988) Private communication. M. E. Muller,
"Some Continuous Monte Carlo Methods for the Dirichlet
Problem," The Annals of Mathematical Statistics, 27(3): 569-589,
1956.
Mikhailov, G. A. (1995) New Monte Carlo Methods With
Estimating Derivatives, V. S. P. Publishers.
Niederreiter, H. (1978) “Quasi-Monte Carlo methods and
pseudo-random numbers," Bull. Amer. Math. Soc., 84: 957–1041.
Rubenstein, M. (1981) Simulation and the Monte Carlo Method,
Wiley-Interscience: New York.
Sabelfeld, K. K. (1991), Monte Carlo Methods in Boundary Value
Problems, Springer-Verlag, Berlin, Heidelberg, New York.
Sabelfeld, K. and N. Mozartova (2009) “Sparsified
Randomization Algorithms for large systems of linear equations
and a new version of the Random Walk on Boundary method,"
Monte Carlo Methods and Applications, 15(3): 257–284.
MCM for PDEs
Bibliography
Bibliography
McKean, H. P. (1988) Private communication. M. E. Muller,
"Some Continuous Monte Carlo Methods for the Dirichlet
Problem," The Annals of Mathematical Statistics, 27(3): 569-589,
1956.
Mikhailov, G. A. (1995) New Monte Carlo Methods With
Estimating Derivatives, V. S. P. Publishers.
Niederreiter, H. (1978) “Quasi-Monte Carlo methods and
pseudo-random numbers," Bull. Amer. Math. Soc., 84: 957–1041.
Rubenstein, M. (1981) Simulation and the Monte Carlo Method,
Wiley-Interscience: New York.
Sabelfeld, K. K. (1991), Monte Carlo Methods in Boundary Value
Problems, Springer-Verlag, Berlin, Heidelberg, New York.
Sabelfeld, K. and N. Mozartova (2009) “Sparsified
Randomization Algorithms for large systems of linear equations
and a new version of the Random Walk on Boundary method,"
Monte Carlo Methods and Applications, 15(3): 257–284.
MCM for PDEs
Bibliography
Bibliography
McKean, H. P. (1988) Private communication. M. E. Muller,
"Some Continuous Monte Carlo Methods for the Dirichlet
Problem," The Annals of Mathematical Statistics, 27(3): 569-589,
1956.
Mikhailov, G. A. (1995) New Monte Carlo Methods With
Estimating Derivatives, V. S. P. Publishers.
Niederreiter, H. (1978) “Quasi-Monte Carlo methods and
pseudo-random numbers," Bull. Amer. Math. Soc., 84: 957–1041.
Rubenstein, M. (1981) Simulation and the Monte Carlo Method,
Wiley-Interscience: New York.
Sabelfeld, K. K. (1991), Monte Carlo Methods in Boundary Value
Problems, Springer-Verlag, Berlin, Heidelberg, New York.
Sabelfeld, K. and N. Mozartova (2009) “Sparsified
Randomization Algorithms for large systems of linear equations
and a new version of the Random Walk on Boundary method,"
Monte Carlo Methods and Applications, 15(3): 257–284.
MCM for PDEs
Bibliography
Bibliography
McKean, H. P. (1988) Private communication. M. E. Muller,
"Some Continuous Monte Carlo Methods for the Dirichlet
Problem," The Annals of Mathematical Statistics, 27(3): 569-589,
1956.
Mikhailov, G. A. (1995) New Monte Carlo Methods With
Estimating Derivatives, V. S. P. Publishers.
Niederreiter, H. (1978) “Quasi-Monte Carlo methods and
pseudo-random numbers," Bull. Amer. Math. Soc., 84: 957–1041.
Rubenstein, M. (1981) Simulation and the Monte Carlo Method,
Wiley-Interscience: New York.
Sabelfeld, K. K. (1991), Monte Carlo Methods in Boundary Value
Problems, Springer-Verlag, Berlin, Heidelberg, New York.
Sabelfeld, K. and N. Mozartova (2009) “Sparsified
Randomization Algorithms for large systems of linear equations
and a new version of the Random Walk on Boundary method,"
Monte Carlo Methods and Applications, 15(3): 257–284.
MCM for PDEs
Bibliography
Bibliography
McKean, H. P. (1988) Private communication. M. E. Muller,
"Some Continuous Monte Carlo Methods for the Dirichlet
Problem," The Annals of Mathematical Statistics, 27(3): 569-589,
1956.
Mikhailov, G. A. (1995) New Monte Carlo Methods With
Estimating Derivatives, V. S. P. Publishers.
Niederreiter, H. (1978) “Quasi-Monte Carlo methods and
pseudo-random numbers," Bull. Amer. Math. Soc., 84: 957–1041.
Rubenstein, M. (1981) Simulation and the Monte Carlo Method,
Wiley-Interscience: New York.
Sabelfeld, K. K. (1991), Monte Carlo Methods in Boundary Value
Problems, Springer-Verlag, Berlin, Heidelberg, New York.
Sabelfeld, K. and N. Mozartova (2009) “Sparsified
Randomization Algorithms for large systems of linear equations
and a new version of the Random Walk on Boundary method,"
Monte Carlo Methods and Applications, 15(3): 257–284.
MCM for PDEs
Bibliography
Bibliography
Sherman, A. S., and C. S. Peskin (1986) “A Monte Carlo method
for scalar reaction diffusion equations”, SIAM
J. Sci. Stat. Comput., 7: 1360–1372.
Sherman, A. S., and C. S. Peskin (1988) “Solving the
Hodgkin-Huxley equations by a random walk method”, SIAM
J. Sci. Stat. Comput., 9: 170–190.
Shreider, Y. A. (1966) The Monte Carlo Method. The Method of
Statistical Trial, Pergamon Press: New York.
Spanier, J. and E. M. Gelbard (1969) Monte Carlo Principles and
Neutron Transport Problems, Addison-Wesley: Reading, MA.
Wasow, W. R. (1952), “A Note on the Inversion of Matrices by
Random Walks," Mathematical Tables and Other Aids to
Computation, 6(38): 78–81.
MCM for PDEs
Bibliography
Bibliography
Sherman, A. S., and C. S. Peskin (1986) “A Monte Carlo method
for scalar reaction diffusion equations”, SIAM
J. Sci. Stat. Comput., 7: 1360–1372.
Sherman, A. S., and C. S. Peskin (1988) “Solving the
Hodgkin-Huxley equations by a random walk method”, SIAM
J. Sci. Stat. Comput., 9: 170–190.
Shreider, Y. A. (1966) The Monte Carlo Method. The Method of
Statistical Trial, Pergamon Press: New York.
Spanier, J. and E. M. Gelbard (1969) Monte Carlo Principles and
Neutron Transport Problems, Addison-Wesley: Reading, MA.
Wasow, W. R. (1952), “A Note on the Inversion of Matrices by
Random Walks," Mathematical Tables and Other Aids to
Computation, 6(38): 78–81.
MCM for PDEs
Bibliography
Bibliography
Sherman, A. S., and C. S. Peskin (1986) “A Monte Carlo method
for scalar reaction diffusion equations”, SIAM
J. Sci. Stat. Comput., 7: 1360–1372.
Sherman, A. S., and C. S. Peskin (1988) “Solving the
Hodgkin-Huxley equations by a random walk method”, SIAM
J. Sci. Stat. Comput., 9: 170–190.
Shreider, Y. A. (1966) The Monte Carlo Method. The Method of
Statistical Trial, Pergamon Press: New York.
Spanier, J. and E. M. Gelbard (1969) Monte Carlo Principles and
Neutron Transport Problems, Addison-Wesley: Reading, MA.
Wasow, W. R. (1952), “A Note on the Inversion of Matrices by
Random Walks," Mathematical Tables and Other Aids to
Computation, 6(38): 78–81.
MCM for PDEs
Bibliography
Bibliography
Sherman, A. S., and C. S. Peskin (1986) “A Monte Carlo method
for scalar reaction diffusion equations”, SIAM
J. Sci. Stat. Comput., 7: 1360–1372.
Sherman, A. S., and C. S. Peskin (1988) “Solving the
Hodgkin-Huxley equations by a random walk method”, SIAM
J. Sci. Stat. Comput., 9: 170–190.
Shreider, Y. A. (1966) The Monte Carlo Method. The Method of
Statistical Trial, Pergamon Press: New York.
Spanier, J. and E. M. Gelbard (1969) Monte Carlo Principles and
Neutron Transport Problems, Addison-Wesley: Reading, MA.
Wasow, W. R. (1952), “A Note on the Inversion of Matrices by
Random Walks," Mathematical Tables and Other Aids to
Computation, 6(38): 78–81.
MCM for PDEs
Bibliography
Bibliography
Sherman, A. S., and C. S. Peskin (1986) “A Monte Carlo method
for scalar reaction diffusion equations”, SIAM
J. Sci. Stat. Comput., 7: 1360–1372.
Sherman, A. S., and C. S. Peskin (1988) “Solving the
Hodgkin-Huxley equations by a random walk method”, SIAM
J. Sci. Stat. Comput., 9: 170–190.
Shreider, Y. A. (1966) The Monte Carlo Method. The Method of
Statistical Trial, Pergamon Press: New York.
Spanier, J. and E. M. Gelbard (1969) Monte Carlo Principles and
Neutron Transport Problems, Addison-Wesley: Reading, MA.
Wasow, W. R. (1952), “A Note on the Inversion of Matrices by
Random Walks," Mathematical Tables and Other Aids to
Computation, 6(38): 78–81.
MCM for PDEs
Bibliography
c
Michael
Mascagni, 2011
Download